This PDF file contains the front matter associated with SPIE Proceedings Volume 12567, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Leukemia diagnosis and therapy planning are both based on classifying peripheral blood images, under a high inter/intra observer variability scenario. In such applications, automatic image processing and classification strategies have obtained outstanding recognition results, however they are fully dependent on the quality of the annotated data. Unlike supervised classification approaches which built upon label-transformations, the herein presented methodology introduces an unsupervised White Blood Cell characterization in the latent space of a Variational Autoencoder (VAE). The latent space is constructed upon 128 parameters from 64 gaussian distributions and then a k-means clustering may retrieve leukemia diagnostic meaningful cell groups. The whole procedure is twofold assessed: 1) evaluation of the 128 dimension VAE latent space for differentiating cells with higher diagnostic value (blast cells) from other peripheral blood components under multiple supervised classification strategies, and 2) quantification of VAE-parameter clustering capacity to unsupervised separation of blast and non-blast cells. Obtained accuracies of each experiment, 0.888 and 0.757 respectively, suggest that the presented strategy successfully characterizes white blood cells and provides a representation space where subtle cell differences can be objectively measured.

Clear cell renal cell carcinoma (ccRCC) is a common cancer and could result in poor prognosis. Understanding individual tumor immune microenvironment (TIME) in ccRCC patients may predict prognosis and response to therapy. In this work, we explore the concept of using radiomic features extracted from computer tomography (CT) imaging to correlate the TIME measurements from multiplex immunohistochemistry (mIHC) analysis. Since CT imaging has long been the standard for evaluation of RCCs, it has the potential to provide noninvasive approximations of the tissue-based mIHC biomarkers. We selected two biomarkers that were grounded by clinical research: PD-L1 expression and CD8⁺PD-1⁺ T cell to CD8⁺ T cell ratio of the tumor epithelium. Then we extracted these two markers from a preliminary set of 52 patients using automated mIHC analysis. We used Random Forest, AdaBoost and ElasticNet to classify each sample as either expressing high or low levels of these markers. We found the radiomic features can correlate tumor epithelium PD-L1 >5%, PD-L1 >10%, and CD8⁺PD1⁺/CD8⁺ >37% with AUROC 0.75, 0.85 and 0.71, respectively.

The muscle coactivation phenomenon (mCoA) has been used to describe motor tasks such as human gait, reflecting the levels of joint stability of the lower limbs in healthy people versus patients with musculoskeletal problems. Nonetheless, scarce literature has described the mCoA evolution in patients with Achilles tendon reconstruction surgery. The aim of this study was to describe the mCoA during human gait in one patient undergoing Achilles tendon reconstruction surgery for acute tendon rupture with a new surgical technique (PARS-Dresden) in two times-frame (12 and 24 weeks). Human gait was recorded on a treadmill at 2.5 km/h. This task was recorded with a 3D motion system synchronized with electromyographic (sEMG) equipment. The sEMG signals were acquired from three portions of the triceps surae (soleus [So], medial [MG], and lateral [LG] gastrocnemius) and tibialis anterior [TA], in both shanks. The mCoA levels (represented as a percentage) were calculated considering twenty gait cycles in muscular pairs TA-MG, TA-LG, and TA-So, in both operated (OL) and non-operated (NOL) limbs. The analysis was estimated with a processing pipeline developed in Python language, considering the Falcone and Winter equation. The results showed a slight increase in the mCoA during gait in some muscular pairs at 24 weeks. The coactivation levels were greater in the TA-MG and TA-So pairs of the OL vs the NOL. Additionally, a higher asymmetry was observed between OL vs NOL at the 12 weeks.

The Leap Motion Controller (LMC) is a hand-held, non-professional motion capture device that is able to evaluate different motor tasks in a clinical setting. In this study, its accuracy was verified, evaluating its performance in the acquisition of movements based on the simulation of the Archimedean spiral test. We present an innovative method for obtaining information on fine motor skills during the spiral drawing test, using an instrumented evaluation platform based on the LMC. First, several experimental conditions were carried out that challenged the performance of the LMC during the simulation of the spiral test, then a case study was conducted to establish the validity of its use to evaluate the spiral drawing test in humans. Due to this research it is possible to extrapolate the spiral drawing test from a two-dimensional space to a three-dimensional space, without restrictions, or contact with objects and in a controlled clinical environment. Allowing to obtain variables of the kinematics of the hand during a spiral drawing task in a three-dimensional space in future.

Due to aging of the population, some studies predict that the burden of Parkinson Disease (PD) will grow substantially in future decades. The rapid increase of PD will place a substantial burden on individuals, society, and health systems. In recent years, a series of works have been published on the use of mobile devices, equipped with sensors, such as accelerometers, gyroscopes and magnetometers to diagnosis and monitor PD outpatients . In this work, the influence of a series of factors on the diagnosis of Parkinson disease were evaluated, using walking activity data obtained from an mPower study. Through constructing several databases, the following factors were evaluated: dependent individual and independent individual approach, input record size, interleaved and non-interleaved data. In addition to these factors, the effect of the complexity of the CNN network on its performance was also evaluated. Databases with large records provided models with better performance in PD diagnosis than databases with small records. CNN's complexity also had a great impact on PD diagnosis performance. In this work, the best results achieved for the independent individual approach and for the dependent individual approach were an AUCROC of 0.511 and 0.861, respectively.

3D tooth segmentation is an important task for digital orthodontics. Several Deep Learning methods have been proposed for automatic tooth segmentation from 3D dental models or intraoral scans. These methods require annotated 3D intraoral scans. Manually annotating 3D intraoral scans is a laborious task. One approach is to devise self-supervision methods to reduce the manual labeling effort. Compared to other types of point cloud data like scene point cloud or shape point cloud data, 3D tooth point cloud data has a very regular structure and a strong shape prior. We look at how much representative information can be learnt from a single 3D intraoral scan. We evaluate this quantitatively with the help of ten different methods of which six are generic point cloud segmentation methods whereas the other four are tooth segmentation specific methods. Surprisingly, we find that with a single 3D intraoral scan training, the Dice score can be as high as 0.86 whereas the full training set gives Dice score of 0.94. We conclude that the segmentation methods can learn a great deal of information from a single 3D tooth point cloud scan under suitable conditions e.g. data augmentation. We are the first to quantitatively evaluate and demonstrate the representation learning capability of Deep Learning methods from a single 3D intraoral scan. This can enable building self-supervision methods for tooth segmentation under extreme data limitation scenario by leveraging the available data to the fullest possible extent.

Transfer learning is a widely used technique in medical imaging and other research fields where a scarcity of available data limit the training of machine learning algorithms. Despite its widespread use and extensive supporting body of research, the specific mechanisms behind transfer learning are not completely understood. In this work, we quantify the effectiveness of transfer learning in medical image classification scenarios for different numbers of training set images. We trained ResNet50, a popular deep learning model used in medical image classification, using two scenarios: 1) applying transfer learning to a pre-trained network and 2) training the same model from scratch (i.e., starting with randomly selected weights). We analyzed the performance of the model under both scenarios as the number of training set images increased from 5,000 to 160,000 medical images. We introduced and evaluated a metric, the transfer learning gap (TLG), to quantify the differences between the two scenarios. The TLG measured the difference in the area under the loss curves (AULCs) when transfer learning was applied and when the model was trained from scratch. Our experiments show that as the training set size increases, the TLG trends to zero, suggesting that the advantage of using transfer learning decreases. The trend in the AULC suggests a training set size where the two scenarios would have equal losses. At this point, the model reaches the same performance regardless of if transfer learning or training from scratch was used. This study is important because it provides a novel metric to understand and quantify the effect of transfer learning.

Segmentation of brain arterio-venous malformations (bAVMs) in 3D rotational angiographies (3DRA) is still an open problem in the literature, with high relevance for clinical practice. While deep learning models have been applied for segmenting the brain vasculature in these images, they have never been used in cases with bAVMs. This is likely caused by the difficulty to obtain sufficiently annotated data to train these approaches. In this paper we introduce a first deep learning model for blood vessel segmentation in 3DRA images of patients with bAVMs. To this end, we densely annotated 5 3DRA volumes of bAVM cases and used these to train two alternative 3DUNet-based architectures with different segmentation objectives. Our results show that the networks reach a comprehensive coverage of relevant structures for bAVM analysis, much better than what is obtained using standard methods. This is promising for achieving a better topological and morphological characterisation of the bAVM structures of interest. Furthermore, the models have the ability to segment venous structures even when missing in the ground truth labelling, which is relevant for planning interventional treatments. Ultimately, these results could be used as more reliable first initial guesses, alleviating the cumbersome task of creating manual labels.

Whole brain tractography is commonly used to study the brain’s white matter fiber pathways, but the large number of streamlines generated - up to one million per brain - can be challenging for large-scale population studies. We propose a robust dimensionality reduction framework for tractography, using a Convolutional Variational Autoencoder (ConvVAE) to learn low-dimensional embeddings from white matter bundles. The resulting embeddings can be used to facilitate downstream tasks such as outlier and abnormality detection, and mapping of disease effects on white matter tracts in individuals or groups. We design experiments to evaluate how well embeddings of different dimensions preserve distances from the original high-dimensional dataset, using distance correlation methods. We find that streamline distances and inter-bundle distances are well preserved in the latent space, with a 6-dimensional optimal embedding space. The generative ConvVAE model allows fast inference on new data, and the smooth latent space enables meaningful decodings that can be used for downstream tasks. We demonstrate the use of a ConvVAE model trained on control subjects’ data to detect structural anomalies in white matter tracts in patients with Alzheimer’s disease (AD). Using ConvVAEs to facilitate population analyses, we identified 6 tracts with statistically significant differences between AD and controls after controlling for age and sex effect, visualizing specific locations along the tracts with high anomalies despite large inter-subject variations in fiber bundle geometry.

Among all the tumors that can affect the brain, gliomas are the most frequent, thus is important to get a correct characterization and delimitation of this malformation to provide the best diagnosis and treatment possible. Nevertheless, there are some issues when dealing with segmenting tumors, it can be a long and tedious labor, which makes it prone to mistakes. To solve those problems, several techniques were proposed, including automatic and semiautomatic segmentation. In this work, we propose the use of a U-net architecture-based deep neural network to automatically realize segmentations of tumors on magnetic resonance brain images obtained from the BRATS 2020 database, which provides T1, T1 contrast enhancement (T1ce), T2, and FLAIR images for each subject. The database has a total of 1476 images distributed in 369 patients, that were shuffled into the training set with 70% of the subjects, and the test set with a percentage of 30%. Our results got a 91.6% DICE value for the validation, from a 91.6% for necrotic core (NET), 91.7% for peritumoral edema (PE), and a 91.4% for enhancing tumor (ET). After the training, we got 55.5%,66.5%, and 68.6% DICE values for NET, PE and ET respectively. We also calculated the whole tumor (WT) segmentation performance, reaching a 78.8% precision and the tumor core (TC) segmentation which reach 75,5% precision.

Machine learning methods have been used for over a decade for staging and subtyping a variety of brain diseases, offering fast and objective methods to classify neurodegenerative diseases such as Alzheimer’s disease (AD). Deep learning models based on convolutional neural networks (CNNs) have also been used to infer dementia severity and predict future clinical decline. Most CNN-based deep learning models use T1-weighted brain MRI scans to identify predictive features for these tasks. In contrast, we examine the added value of diffusion-weighted MRI (dMRI) - a variant of MRI, sensitive to microstructural tissue properties - as an additional input in CNN-based models of dementia severity. dMRI is sensitive to microstructural brain abnormalities not evident on standard anatomical MRI. By training CNNs on combined anatomical and diffusion MRI, we hypothesize that we could boost performance when predicting widely-used clinical assessments of dementia severity, such as individuals’ scores on the ADAS11, ADAS13, and MMSE (mini-mental state exam) clinical scales. For benchmarking, we evaluate CNNs that use T1-weighted MRI and dMRI to estimate “brain age” - the task of predicting a person’s chronological age from their neuroimaging data. To assess which dMRI-derived maps were most beneficial, we computed DWI-derived diffusion tensor imaging (DTI) maps of mean and radial diffusivity (MD/RD), axial diffusivity (AD) and fractional anisotropy (FA) for 1198 elderly subjects (age: 74.35 +/- 7.74 yrs.; 600 F/598 M, with a distribution of 636 CN/421 MCI/141 AD) from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). We tested both 2D Slice CNN and 3D CNN neural network models for the above predictive tasks. Our results suggest that for at least some deep learning architectures, diffusion-weighted MRI may enhance performance for several AD-relevant deep learning tasks relative to using T1-weighted images alone.

Lesion appearance is a crucial clue for medical providers to distinguish referable diabetic retinopathy (rDR) from non-referable DR. Most existing large-scale DR datasets contain only image-level labels rather than pixel-based annotations. This motivates us to develop algorithms to classify rDR and segment lesions via image-level labels. This paper leverages self-supervised equivariant learning and attention-based multi-instance learning (MIL) to tackle this problem. MIL is an effective strategy to differentiate positive and negative instances, helping us discard background regions (negative instances) while localizing lesion regions (positive ones). However, MIL only provides coarse lesion localization and cannot distinguish lesions located across adjacent patches. Conversely, a self-supervised equivariant attention mechanism (SEAM) generates a segmentation-level class activation map (CAM) that can guide patch extraction of lesions more accurately. Our work aims at integrating both methods to improve rDR classification accuracy. We conduct extensive validation experiments on the Eyepacs dataset, achieving an area under the receiver operating characteristic curve (AU ROC) of 0.958, outperforming current state-of-the-art algorithms.

Ascites is often regarded as the hallmark of advanced ovarian cancer, which is the most lethal gynecologic malignancy. Ascites segmentation contributes to track the progress of ovarian cancer development by providing accurate ascites measurement, which can effectively guide subsequent treatment and potentially reduce the mortality. Segmentation of ascites is challenging due to the presence of iso-intense fluids such as bile, urine, etc., near the ascites region. In this work we propose a novel 3D U-Net segmentation method called body location embedded U-Net (BLE-U-Net) that integrates anatomical location information with the segmentation process. BLE-U-Net incorporates body part regression to predict the approximate anatomical location of each CT slice along the z- axis. The regression scores are discretized to indicate different body regions and embedded into a modified 3D U-Net to improve the ascites segmentation. Twenty contrast-enhanced body CT scans were used to evaluate the proposed method. Dice coefficients of 38 ±10 and 65 ±06 were achieved for a conventional 3D U-Net and BLE-U-Net, respectively (with t-test p <0.05). Volumes of segmented ascites were 0.51±0.74 and 0.57±0.85 liters for each method where the ground-truth volume was 0.58±0.84 liters. These results suggest that the embedded location information is the key factor to improve the ascites segmentation, which could potentially benefit ovarian cancer diagnosis and treatment.

t-Distributed Stochastic Neighbor Embedding (t-SNE) and k-means have been increasingly utilized for dimension reduction and graphical illustration in medical imaging (e.g., CT) informatics. Mapping a grid network onto a slide is a prerequisite for implementing cluster analysis. Traditionally, the performance of cluster analysis is driven by hyperparameters, however, grid size which also affects performance is often set arbitrarily. In this study, we evaluated the effect of varying grid sizes, perplexity and learning rate hyperparameters for unsupervised clustering using CT images of renal masses. We investigated the impact of grid size to cluster analysis. The number of clusters was determined by Gap-statistics. The grid size selections were 2x2, 4x4, 5x5, and 8x8. The results showed that the number of output clusters increased with decreasing grid sizes from 8x8 to 4x4. However, when grid size reached 2x2, the model yielded the same cluster number as 8x8. This finding was consistent across different hyperparameter settings. Additional analyses were conducted to understand the nesting structure between the cluster membership (the mutually exclusive cluster number assigned to each grid in a cluster analysis) from large (8x8) grid and small (2x2) grid, although both grid size selections yielded the same number of clusters. We report that the cluster membership between large grid and small grid is only partially overlaid. This suggests that additional pattern/information is detected by using the small grid. In conclusion, the grid size should be treated as another hyperparameter when using unsupervised clustering methods for pattern recognition in medical imaging analysis.

In this prospective study, forty patients with solid renal masses who underwent contrast-enhanced ultrasound (CEUS) examinations were selected. Using the ImageJ software, renal masses and adjacent normal tissue were manually segmented from CEUS cine exams obtained using the built-in RS85 Samsung scanner software. For the radiomics analysis, one frame representing precontrast, early, peak, and delay enhancement phase were selected post segmentation from each CEUS clip. From each region of interest (ROI) within a tumor tissue normalized renal mass, 112 radiomic metrics were extracted using custom Matlab® code. For the time-intensity curve (TIC) analysis, the segmented ROIs were plotted as a function of time, and the data were fit to a washout curve. From these time-signal intensity curves, perfusion quantitative parameters, were generated. Wilcoxon rank sum test or univariate independent t-test depending on data normality were used for descriptive analyses. Agreement was analyzed using Kappa statistic. Of the 40 solid masses, 31 (77.5%) were malignant, 9 (22.5%) were benign based on histopathology. Excellent agreement was found between histopathological confirmation and visual assessment based on CEUS in discriminating solid renal masses into benign vs. malignant categories (κ=0.89 95% confidence interval (CI): (0.77,1)). The total agreement between the two was 92.5%. The sensitivity and specificity of CEUS-based visual assessment was found to be 100% and 66.7%, respectively. Quantitative analysis revealed TIC metrics revealed statistically significant differences between the malignant and benign groups and between clear cell renal cell carcinoma (ccRCC) and papillary renal cell carcinoma (pRCC) subtypes. The study shows excellent agreement between visual assessment and histopathology, but with the room to improve in specificity.

In this IRB approved retrospective study 41 women with biopsy-proven invasive breast cancers (IBC) were imaged using contrast-enhanced mammography (CEM), prior to any treatment. Size-matched regions of interest (ROIs) were manually contoured by an experienced breast radiologist on the CEM capturing the breast lesion and breast parenchymal enhancement (BPE), respectively. Radiomics analysis was performed using LifEx software and 109 radiomics metrics spanning 6 different texture families were extracted from each ROI. Predictive models of lesion malignancy were developed using multiple classifiers and used to subclassify breast cancers based on their hormone receptor status. The 10- fold cross validation was used to construct the decision classifier and performance was assessed. CEM radiomics models based on Random Forest, Real Adaboost, and ElasticNet classifiers achieved an AUC of 0.83, 0.82 and 0.74, respectively in discriminating malignant breast lesions from varying amounts of BPE. Accounting for the varying levels of BPE, revealed a reduction in AUC-based prediction of lesion vs. BPE as the qualitative assessment of BPE increased from minimal to moderate (AUCs of 0.89 vs 0.74). Further analyses of the IBC based on their hormone receptor status showed that triple negative breast lesions showed statistically significant differences in multiple radiomics metrics compared to ER+ PR+ HER2- and HER2+. The predicted probability of the radiomics model was significantly different across three receptor-based subtypes and between high and low nuclear grade breast cancers. CEM Radiomics demonstrated good discrimination (AUC>0.8) of malignant breast lesions despite varying BPE levels and supports breast lesion subtyping.

Morphological metrics such as fractal dimension (FD) have shown value as diagnostic and prognostic markers in diverse cancers. A lack of procedural consensus on fractal techniques may lead to a non-generalization of results across different studies. This study reports variations of Computed Tomography (CT) derived FD renal masses across different fractal analysis implementations. The Fraclac grayscale pixel size 512x512 pixel setting Area Under Curve (AUC) showed the highest AUC value (0.59) among all pixel settings in classifying clear cell renal cell carcinoma (ccRCC) vs. Oncocytoma and liquid poor angiomyolipoma (AML). Similarly, for the multiphase analysis, we also explored MATLAB grayscale pixel sizes from 7x7 to 256x256 pixels. Results showed that the 64x64 pixel setting had the highest AUC of 0.60-0.72 for ccRCC vs. Oncocytoma and AML and AUC of 0.58-0.69 for chromophobe renal cell carcinoma (RCC) vs Oncocytoma.

Currently, there are many methods for processing diffusion MRI (dMRI) tractography data, with the aim to identify the main white matter connections. However, methods like fiber clustering lack ground truth, making the evaluation of the effectiveness of different clustering algorithms problematic. An alternative to evaluate the performance and test the efficacy of these algorithms is to use simulated fiber datasets. Nevertheless, the simulation of this data is not trivial due to brain fibers’ irregular and complex shape. Although many fiber bundle simulators exist, they have been developed for other purposes, such as validating tractography algorithms or local diffusion models. In addition, these simulators usually use simple fiber bundle configurations without considering complex bundle shapes. With this in mind, the main goal of this work is to implement a simulator of brain fiber bundles based on exponential curves for validating fiber clustering methods. This representation uses bundle centroids and shape parameters to obtain a more realistic appearance of the fascicles. The simulator was validated using a deep white matter fiber bundle atlas, obtaining a good percentage of intersection between the original and simulated bundles, of up to 82%. Furthermore, we used groups of simulated bundles for the whole brain to evaluate the performance of a fiber clustering algorithm (QuickBundles) when using different distance thresholds, showing the utility of the proposed simulator.

Parkinson’s disease is a prevalent neurodegenerative disorder affecting millions of people worldwide. It is caused by degeneration of dopaminergic neurons of the substantia nigra. This has given rise to imaging-based biomarker discovery studies using neuromelanin-sensitive MRI. In this paper, we propose a normative template constructed using a combination of structural and neuromelanin-sensitive MRI that can be used in biomarker studies of Parkinson’s disease. In contrast to previous template construction approaches, the proposed combination of images allows to impose further constraints on the registrations used to create the template. Our experiments demonstrate that the proposed template construction pipeline yields a consistent localisation of the substantia nigra and other brainstem structures and that the registrations used for this purpose yield smooth and invertible transformations.

This paper presents a new cortical parcellation method based on group-wise connectivity and hierarchical clustering. A preliminary sub-parcellation is performed using intra-subject and inter-subject fiber clustering to obtain representative bundles among subjects with similar shapes and trajectories. The sub-parcellation is obtained by intersecting fiber clusters with cortical meshes. Next, mean connectivity and mean overlap matrices are computed over the sub-parcels to obtain spatial and connectivity information. To hierarchize the information, we propose to weight both matrices, to obtain an affinity graph, and then a dendrogram to merge or divide parcels by their hierarchy. Finally, to obtain homogeneous parcels, the method computes morphological operations. By selecting a different number of clusters over the dendrogram, the method obtains a different number of parcels and a variation in the resulting parcel sizes, depending on the parameters used. We computed the coefficient of variation (CV ) of the parcel size to evaluate the homogeneity of the parcels. Preliminary results suggest that the use of representative clusters and the integration of sub-parcel overlap and connectivity strength provide useful information to generate cortical parcellations at different levels of granularity. Even results are preliminary, this novel method allows researchers to add group-wise connectivity strength and spatial information for the construction of diffusion-based parcellations. Future work will include a detailed analysis of parameters, such as the matrix weights and the number of sub-parcel clusters, and the generation of hierarchical parcellations to improve the insight into the cortex subdivision and hierarchy among parcels.

Corpus Callosum (CC) segmentation is required when the analysis from this structure is desirable. Many of these studies require the CC segmentation on diffusion tensor images (DTI). However, few methods perform segmentation directly in the DTI. Segmenting on DTI makes it possible to disregard the registration step after segmenting on T1 images. This work studies the possibility of improving automated segmentation of the CC using silver standard annotations. With incomplete silver standard annotations, limited to 5 or 7 central slices, experiments performed throughout this work were done to compare methods of pre-training and fine tuning in an attempt to translate silver standard knowledge to improved performance in 3D CC segmentation. Experiments include 3D and 2D U-Net as deep learning architectures. Results point to central limited silver standard annotations not being useful for improving the performance in gold standard 3D annotations. Our best method involved training a 3D U-Net with gold standards and post processing, achieving a 3D Dice of 83.33 Dice, surpassing 2D U-Net.

The description of the superficial white matter (SWM) functional and structural organization is still an unachieved task. In particular, their shape has not been assessed in detail using diffusion Magnetic Resonance Imaging (dMRI) tractography. This work aims to characterize the different shapes of the short-range association connections present in an SWM multi-subject bundle atlas derived from probabilistic dMRI tractography datasets. First, we calculated a representative centroid shape for each atlas bundle. Next, we computed a distance matrix that encodes the similarity between every pair of centroids. For the distance matrix computation, centroids were first aligned using a streamline-based registration, reducing the 3D spatial separation effect and allowing us to focus only on shape differences. Then, we applied a hierarchical clustering algorithm over the affinity graph derived from the distance matrix. As a result, we obtained ten classes with distinctive shapes, ranging from a straight line form to U and C arrangements. The most predominant shapes were: (i) short open U, (ii) short closed U, and (iii) short C. Moreover, we used the shape information to filter out noisy streamlines in the atlas bundles and applied an automatic segmentation algorithm to 25 subjects of the HCP database. Our results show that the filtering steps help to segment more dense bundles with fewer outliers, improving the identification of the brain’s short fibers.

Brain tumor image segmentation is a challenging research topic in which deep-learning models have presented the best results. However, the traditional way of training those models from many pre-annotated images leaves several unanswered questions. Hence methodologies, such as Feature Learning from Image Markers (FLIM), have involved an expert in the learning loop to reduce human effort in data annotation and build models sufficiently deep for a given problem. FLIM has been successfully used to create encoders, estimating the filters of all convolutional layers from patches centered at marker voxels. In this work, we present Multi-Step (MS) FLIM – a user-assisted approach to estimating and selecting the most relevant filters from multiple FLIM executions. MS-FLIM is used only for the first convolutional layer, and the results already indicate improvement over FLIM. For evaluation, we build a simple U-shaped encoder-decoder network, named sU-Net, for glioblastoma segmentation using T1Gd and FLAIR MRI scans, varying the encoder’s training method, using FLIM, MS-FLIM, and backpropagation algorithm. Also, we compared these sU-Nets with two State-Of-The-Art (SOTA) deep-learning models using two datasets. The results show that the sU-Net based on MS-FLIM outperforms the other training methods and achieves effectiveness within the standard deviations of the SOTA models.

Deep learning is currently the state-of-the-art for automated detection of referable diabetic retinopathy (DR) from color fundus photographs (CFP). While the general interest is put on improving results through methodological innovations, it is not clear how good these approaches perform compared to standard deep classification models trained with the appropriate settings. In this paper we propose to model a strong baseline for this task based on a simple and standard ResNet-18 architecture. To this end, we built on top of prior art by training the model with a standard preprocessing strategy but using images from several public sources and an empirically calibrated data augmentation setting. To evaluate its performance, we covered multiple clinically relevant perspectives, including image and patient level DR screening, discriminating responses by input quality and DR grade, assessing model uncertainties and analyzing its results in a qualitative manner. With no other methodological innovation than a carefully designed training, our ResNet model achieved an AUC = 0.955 (0.953 - 0.956) on a combined test set of 61007 test images from different public datasets, which is in line or even better than what other more complex deep learning models reported in the literature. Similar AUC values were obtained in 480 images from two separate in-house databases specially prepared for this study, which emphasizes its generalization ability. This confirms that standard networks can still be strong baselines for this task if properly trained.

Despite advances in deep learning methods aimed at cardiac ventricular segmentation, most algorithms have drawbacks due to low prediction accuracy with images from different MR scans to those trained. It leads to a process that requires time-consuming correction by technicians or specialists. The time in this process is significant mainly due to the large number of image sets to be processed. The lack of description of the algorithms has not allowed repeatability, while commercial software is difficult to access for clinical use or research. However, in cardiac segmentation research, several solutions have already been proposed. This paper presents an opensource cardiac functionality segmentation and evaluation framework, which contemplates a diverse database for network training, a multi domain network architecture that allows model generalization, and pre-and postprocessing algorithms that improve prediction results. The prediction evaluation of the framework shows that Ventseg is 3.66% superior to the trained model and the similarity percentages in the tested MR scores are over 84%. On the other hand, the inter-observer variability analysis, with anonymized data, shows in the different metrics that Ventseg is on par with cardiac segmentation specialists. Finally, the efficiency calculated in an intra-observer test indicates that our framework reduces manual segmentation time by approximately 80%.

Automated optic disc (OD) and optic cup (OC) segmentation in fundus images is relevant to efficiently measure the vertical cup-to-disc ratio (vCDR), a biomarker commonly used in ophthalmology to determine the degree of glaucomatous optic neuropathy. In general this is solved using coarse-to-fine deep learning algorithms in which a first stage approximates the OD and a second one uses a crop of this area to predict OD/OC masks. While this approach is widely applied in the literature, there are no studies analyzing its real contribution to the results. In this paper we present a comprehensive analysis of different coarse-to-fine designs for OD/OC segmentation using 5 public databases, both from a standard segmentation perspective and for estimating the vCDR for glaucoma assessment. Our analysis shows that these algorithms not necessarily outperfom standard multi-class single-stage models, especially when these are learned from sufficiently large and diverse training sets. Furthermore, we noticed that the coarse stage achieves better OD segmentation results than the fine one, and that providing OD supervision to the second stage is essential to ensure accurate OC masks. Moreover, both the single-stage and two-stage models trained on a multi-dataset setting showed results close to other state-of-the-art alternatives in REFUGE and DRISHTI. Finally, we evaluated the models for vCDR prediction in comparison with six ophthalmologists on a subset of AIROGS images, to understand them in the context of inter-observer variability. We noticed that vCDR estimates recovered both from single-stage and coarse-to-fine models can obtain good glaucoma detection results even when they are not highly correlated with manual measurements from experts. To ensure the reproducibility of our study, our results, our multi-expert dataset and further implementation details are made publicly available at https://github.com/eugeniaMoris/sipaim-2022-coarse-to-fine.

Alzheimer’s disease (AD) is a neurodegenerative disease that induces cognitive and functional decline in the elderly. Detecting AD in its early, pre-dementia stages (i.e., during mild cognitive impairment, MCI) is challenging as the rate of atrophy varies among brain regions. This progressive degeneration, nevertheless, potentially can be identified in brain imaging by examining changes in texture descriptors. Textures in images are thought to reveal underlying patterns of structural change in brain tissue. In addition, careful examination of a panel of texture descriptors may allow for improved prediction of disease stages. This study evaluates four common groups of texture descriptors (GLCM—gray-level co-occurrence matrix, GLRLM—gray-level run-length matrix, GLSZM— gray-level size zone matrix, and NGTDM—neighbouring gray tone difference matrix) across 85 anatomical brain regions in three clinical populations (HN – healthy normal, MCI, and AD). Specifically, using statistical hypothesis validation at the regional level, differences in texture descriptors were analyzed in three comparative scenarios (HN vs MCI, HN vs AD, and MCI vs AD) and multiple comparisons were corrected using false discovery rate (FDR). From sixty-one examined texture descriptors, only six (9.8%) were found to be statistically significant (p < 0.05) in all three comparative scenarios after FDR correction – five were GLRLM features and one was GLCM. Our results highlighted significant changes in texture descriptors among HN, MCI, and AD populations, and they align with existing biological and computational findings in the literature.

Edinburgh Postpartum Depression (EPDS) and Breastfeeding Self-Efficacy (BSES) scales are standardized questionnaires to screen for postpartum depression and breastfeeding performance self-perception. On the other hand, Natural Language Processing (NLP) is a machine learning technique that analyses the human language to extract relevant and computer-interpretable information. In this work we proposed the application of an NLP toolchain that includes a typical preprocessing stage and the probabilistic topic modeling performed through the Latent Dirichlet Allocation (LDA) to find out the two most relevant topics within each of six study groups (low, medium, and high scores of BSES and EPDS). Each topic LDA-modeled consisted of 30-word/terms (tokens) which are organized in Venn diagrams, contrasting the mutually exclusive tokens within the low and high scores on each scale. Coherence and log-Perplexity topic modeling performance metrics, were computed. We found that LDA-models have distinguishable tokens between low and high scores of the BSES and EPDS. However, the most remarkable findings were two subset of tokens, one related to newborn care and another to newborn intake, respectively correlated to low and high postpartum depression risk according to EPDS.

Data annotation is often a prerequisite for applying deep learning to medical image segmentation. It is a tedious process that requires substantial guidance from experienced physicians. Adipose tissue labeling on CT scans is particularly time-consuming because adipose tissue is present throughout the entire body. One possible solution is to create inaccurate annotations from conventional (non-deep learning) adipose tissue segmentation methods. This work demonstrates the development of a deep learning model directly from these inaccurate annotations. The model is a multi-scale 3D residual U-Net where the encoder path is composed of residual blocks and the decoder path fuses multi-scale feature maps from different layers of decoder blocks. The training set consisted of 101 patients and the testing set consisted of 14 patients. Ten patients with anasarca were purposely added to the testing dataset as a stress test to evaluate model generality. Anasarca is a medical condition that leads to the generalized accumulation of edema within subcutaneous adipose tissue. Edema creates heterogeneity inside the adipose tissue which is absent in the training data. In comparison with a baseline method of manual annotations, the Dice coefficient improved significantly from 73.4 ± 14.1% to 80.2 ± 7.1% (p < 0.05). The model trained on inaccurate annotations improved the accuracy of adipose tissue segmentation by 7% without the need for any manual annotation.

Telemedicine is one of the most important technologies and services nowadays to overcome face-to-face consultation or intervention, especially, in rural areas. Teleultrasound monitoring is a service that has been useful for lung, thyroid or liver monitoring. In particular, one of the most promising protocols to attend this type of monitoring is sending asynchronous acquired ultrasound videos following the volume sweep imaging protocol; however, one of the main limitations is the poor internet bandwidth that does not allow sending files with a large amount of information. For this reason, in this work, it was proposed to send a single image that includes all the information of the acquisition using panoramic reconstruction of the ultrasound video. To perform the experiments, an ultrasound phantom was used with the commercially available ultrasound scanner with a linear transducer at different acquisition depths of 2.5, 3.7, 4.9, 6.1 and 7.4 cm. For the reconstructions, three phases were used detection of the acquisition moment, by means of optical flow; image processing for the selection of the region of interest and intensity adjustment; and panoramic reconstruction determining the displacement between frames. It was observed that this was successful for the deeper and higher contrast acquisitions.

Diabetes is a high-burden non-communicable disease affecting more than 532 million people worldwide and resulting in a range of life-threatening comorbidities. Pre-identifying high-risk individuals and applying preventive actions will likely reduce the prevalence and health consequences of diabetes. Under this context, we developed and evaluated the first predictive model of diabetes risk that combines both electrocardiography (ECG) and exposome predictors. A comprehensive list of ECG signals and exposome variables were extracted from the UK Biobank, then used to build and compare a set of machine learning models for diabetes risk prediction. Random Forest combining ECGs and exposome variables achieved an 0.82 ± 0.03 AUC when predicting diabetes risk. This integrative model outperformed separate models based on exposome factors or ECG signals alone. These preliminary results indicate the potential of low-cost machine learning models trained from ECG and exposome data to predict diabetes years before its onset.

Detecting and extracting findings in a radiological report is crucial for text mining tasks in several applications. In this case, a labeled process for the image associated with the radiological report in mammography and Spanish context for a computer vision model is required. This paper shows the methodology and process generated for this goal. This paper presents a Named Entity Recognition (NER) approach based on a transformer deep learning model, using a labeled corpus and fine-tuning process to find three concepts that compose a typical finding in a mammographic radiological report: laterality, location, and the finding. We add another concept in the labeled process, the negation, necessary to identify falses positive inside the text that writes the radiologist. Our model achieves an F1 score of 88.24% classifying the three principal concepts for a finding, product of the labeled and fine-tuning process. The results presented here will be used as input for future training work on a computer vision model.

The beamforming technique has been successfully used for real-time ultrasound imaging and applications. Traditionally, the most commercially available ultrasound systems still implement standard Delay-and-Sum (DAS) beamforming for B-mode imaging. This technique performs the time delay and coherent summation of ultrasonic radiofrequency (RF) echoes received by individual transducer elements to align the backscattered signals at the focal point. However, the transducer aperture size and system operating frequency limit the image resolution and contrast achievable with DAS. For this reason, new methods based on adaptive beamforming algorithms, such as Minimum Variance (MV), have been studied to improve the quality of the signal received by the transducer and reduce the effects of noise and interference. This work compares a B-mode ultrasound image generated by the DAS technique and the MV combined with DAS beamforming using Field II acoustic field simulation software. A simulated phantom with 18 targets, separated into three groups, and surrounded by a uniform background, was created. For qualitative analysis, two-dimensional and three-dimensional images simulated using DAS and MV beamformers are presented. The quantitative analyses were employed to compare the performance of the MV over the DAS beamforming using axial and lateral full width at half maximum (FWHM) and geometric distortion ratio (GDR) measurements of the central target group. According to those metrics, no significant changes were observed regarding the axial FWHM. However, the MV method considerably reduced the lateral FWHM by more than 40%, with a minimum GDR of 37%.

Lung ultrasound imaging allows the detection and evaluation of the lung damage generated by COVID-19. However, several infrastructure and logistical limitations prevent them from being carried out in isolated and remote areas. In this work, a system for the acquisition of medical images through asynchronous tele-ultrasounds was developed. The system is based on a graphical user interface, which records the three video cameras, the ultrasound image and the accelerometer simultaneously. The interface was developed according to the Volume Sweep Imaging acquisition protocol. The translational and rotational movement of the transducer are tracked and monitored by the accelerometer and the position of the transducer is obtained from the images acquired by the three video cameras. The results show a correct functioning of the system overall, being viable to be implemented for data acquisition and calculation of error, although in order to validate the error calculation there is still more research to be done.

Previous work has shown the potential of deep learning to predict renal obstruction using kidney ultrasound images. However, these image-based classifiers have been trained with the goal of single-visit inference in mind. We compare methods from video action recognition (i.e. convolutional pooling, LSTM, TSM) to adapt single-visit convolutional models to handle multiple visit inference. We demonstrate that incorporating images from a patient’s past hospital visits provides only a small benefit for the prediction of obstructive hydronephrosis. Therefore, inclusion of prior ultrasounds is beneficial but prediction based on the latest ultrasound is sufficient for patient risk stratification.

The lung is vulnerable to suffering from different diseases, for its diagnosis one of the imaging technologies is ultrasound; however, it requires an expert to perform the acquisition and interpretation of the same. In Peru, many rural areas have little technology and untrained personnel, so one solution is to divide the acquisition and diagnosis. In this way, the radiologist only receives the video and performs the diagnosis through it and the local personnel is trained to perform the Volume Sweep Imaging protocol; however, there are some mistakes during its performance which are detected lately. To solve this problem, the paper proposes the use of three different algorithms: Adaptive Threshold, Threshold + non-consecutive minimum distance analysis and Derivative Analysis + Peak Detection, which show a better accuracy in the longitudinal direction of 77.96%; nevertheless, their work principle needs to be reformulated for transverse acquisition, where accuracy is less than 50%. This result show is possible to validate the protocol for lung ultrasound and give feedback to the locally trained physicians.

In this Institutional Review Board (IRB) approved, Health Insurance Portability and Accountability Act (HIPAA) compliant, prospective study, uncompressed envelope data (RF data) were collected from 100 patients with focal renal masses using an RS80A ultrasound scanner with B-mode and CEUS. By summing and averaging the Nakagami images formed using sliding windows, we use the average ‘m’ to stratify manually segmented masses, using data from both the B-mode and CEUS scans. Wilcoxon rank sum test using an alpha value of 0.05 was used detect differences between the groups. Logistic regression was used for classification and the area under the receiver operator curve (AUC) was used to assess performance. Among the 100 masses, 40 were benign, 37 were malignant based on histopathology, and 23 were radiologically and clinically presumed malignant but with no pathological proof at the time of data analysis. Univariate analyses showed significant (p<0.01) differences between the benign and non-benign masses on both B-mode and CEUS, with non-benign masses having smaller ‘m’. Predictive models constructed using Nakagami parameters extracted from Bmode and CEUS-based RF scans showed an AUC of 0.67 95% CI: (0.56, 0.78) and 0.61 95% CI: (0.5, 0.73), respectively for discriminating benign from non-benign renal masses. The concordance between the two assessments was 95%. We present a framework for characterizing images using speckle textural properties, for example Nakagami analysis, to aid in objective tissue characterization using ultrasound.

The COVID-19 pandemic has challenged many of the healthcare systems around the world. Many patients who have been hospitalized due to this disease develop lung damage. In low and middle-income countries, people living in rural and remote areas have very limited access to adequate health care. Ultrasound is a safe, portable and accessible alternative; however, it has limitations such as being operator-dependent and requiring a trained professional. The use of lung ultrasound volume sweep imaging is a potential solution for this lack of physicians. In order to support this protocol, image processing together with machine learning is a potential methodology for an automatic lung damage screening system. In this paper we present an automatic detection of lung ultrasound artifacts using a Deep Neural Network, identifying clinical relevant artifacts such as pleural and A-lines contained in the ultrasound examination taken as part of the clinical screening in patients with suspected lung damage. The model achieved encouraging preliminary results such as sensitivity of 94% , specificity of 81%, and accuracy of 89% to identify the presence of A-lines. Finally, the present study could result in an alternative solution for an operator-independent lung damage screening in rural areas, leading to the integration of AI-based technology as a complementary tool for healthcare professionals.

Dysphagia is defined as difficulty moving food from the mouth to the stomach. The confirmatory test for the diagnosis of this condition is videofluoroscopy (VFSS) being highly invasive since it exposes the patient to ionizing radiation. Surface electromyography (sEMG) records the electrical activity of the muscles in a non-invasive way, so it can offer an alternative to evaluate swallowing and can be applied in dysphagia. We analyzed sEMG signals from 5 patients who were simultaneously acquired with VFSS. Yogurt and water in volumes of 5, 10, and 20 ml, and 3 g of cracker were administered. Masseter, suprahyoid, and infrahyoid muscle groups were studied bilaterally. An algorithm for automatic detection of the onset and offset was developed. The degree of overlap between the automatic segmentation and visual segmentation was calculated for each muscle group. The average activation time of these muscle groups was compared and plotted concerning the bolus pass through the reference points calculated by VFSS: mandibular line (LM), vallecula (V), and cricopharyngeal muscle (part of upper esophageal sphincter). It was found a higher degree of overlap for the larger volumes (20 ml) and in the following order: masseter, suprahyoid, and infrahyoid. The results show a previous activation in most of the muscle groups before LM and a deactivation when the food passes through the cricopharyngeal sphincter (EC). This study showed that the implemented algorithm served to know the time of muscle activity and the sequence of muscle groups involved in swallowing by using sEMG validated with VFSS.

Voice inverse filtering analysis comprises different methods for the non-invasive estimation of glottal airflow from a speech signal, thus bringing forth relevant information about the vocal function and acoustic excitation during voiced phonation. Most inverse filtering strategies consider a parametric source-filter model of phonation and variants of linear prediction to adjust the model coefficients. However, classical linear prediction is susceptible to impulse-like acoustic excitations produced by abrupt glottal closures. Robust alternatives have been proposed that apply a time-domain weighting function to de-emphasize the detrimental contribution of the impulse-like glottal events. The present study introduces the maximum correntropy criterion-based linear prediction for voice inverse filtering. This method takes advantage of the correntropy –a non-linear localized similarity measure inherently insensitive to outliers– to implement a robust weighted linear prediction, where the weighting function is adjusted iteratively through a speech-data-guided optimization scheme. Simulations show that the proposed method naturally overweights samples in the glottal closed phase, where the phonation model is more accurate, without being necessary any prior information about the closure instants. It is further shown that maximum correntropy criterion-based linear prediction improves inverse filtering analysis in terms of the smoothness of estimated glottal waveforms, and the spectral relevance of the vocal tract filter.

Phonetically balanced texts are used to study different voice and speech characteristics. In the context of clinical work and research, these texts provide a standard for quantifying perceptual, acoustic, or aerodynamic assessments. Recent modeling efforts are being devoted to describing long-term speech behaviors based on a collection of sustained phonemes. However, comprehensive descriptions of phoneme distributions representative of connected speech are not readily available. Thus, the present study introduces a method to estimate phoneme distributions using text data mining, as an alternative to existing power law methods. The procedure used for the decomposition of texts into phonemes, the estimation of the phonetic distributions and the comparisons between different texts, conversational speech, and standard reading passages are discussed. The results are presented using histograms and R-squared determination coefficients for the case of the English language, although the approach can be easily applied for other languages. A discussion of the proposed method, results, and limitations is presented.

Current methods for monitoring laryngeal muscle function include techniques such as intramuscular electromyography, external laryngeal palpation, and laryngeal endoscopy. Although these methods have provided much information about muscle activation and function during voice production, they are invasive, uncomfortable, and subjective. The objective of this work is to explore the use of the high-density electromyography (HDsEMG) as a non-invasive alternative that can potentially provide objective information on the activity of the laryngeal muscles during speech. with a focus on the cricothyroid muscle (CT). From this set of signals, it is possible to decompose the electromyography signal, providing indirect information on the spatial recruitment and firing rates of motor units (MU) within the muscle. It is hypothesized that the use of MU firing rate and recruitment will allow for a better estimation of muscle activation compared to traditional methods. A high-density wireless HDsEMG equipment (Sessantaquattro, OT Bioelettronica) is used with a 64-channel electrode grid, which are centered on the CT muscle. Preliminary results of a case study illustrated that it was possible to obtain the rates and firing trains of 4 motor units. Future work will explore how HDsEMG applied to the larynx has the potential to improve diagnostic and therapeutic follow-up of pathologies of laryngeal function.

The brain has the ability to evaluate unattended social information, such as facial expressions, and reassign attentional resources to specific relevant features. Two neuronal mechanisms could account for such facial emotional processing: one slow and accurate system that can be measured around 170 ms and one fast and imprecise system that is triggered around 90 ms which could support early negative emotional processing for automatic/unattended and peripheral stimulation. Evidence that these mechanisms exist for positive affective processing is scarce. The present study investigated the neural correlates of unattended negative and positive emotional processing using the rapid presentation of unilateral and bilateral peripheral facial expressions. Hence, we measured the electrophysiological correlates of unattended fear, happy and neutral faces presented in the left and right hemifields of neurotypical individuals using a frequency tagging paradigm and electroencephalography. Frequency stimulations of 5.8 Hz and 11 Hz were chosen to induce Steady-State Visual Evoked Potential (SSVEP) occurring at 170 ms and 90 ms, respectively. The SSVEP amplitudes showed that unattended positive and negative information in the periphery was processed at early stages and increased the brain's response to attended salient emotional stimuli in posterior visual regions. These results suggest that emotional stimuli presented outside the attentional focus elicit increased brain activity, particularly in posterior regions which could be altered in disorders of social recognition.

Parkinson’s disease (PD) is a neurodegenerative brain disorder that mainly affects the elderly population and currently has no cure, only palliative treatments. Recently, spinal cord stimulation (SCS) showed great promise in the treatment of certain PD symptoms. In this semi-invasive technique, pulses of current are delivered to the spinal cord via electrodes implanted in the epidural space. However, the effects of SCS in the brain are poorly understood. In this work, we developed a method for detection of electrophysiological events in the brain evoked by SCS pulses, and analyzed the latency of such events under different experimental conditions. We performed in vivo recordings of local field potentials using 64 microelectrodes implanted in different brain areas of a rat model of PD, and during active and inactive states, and with and without administration of L-dopa. The signals were pre-processed and divided into windows centered around the stimulation pulses, in which we detected the evoked events using two approaches: the Hampel identifier and the wavelet transform. Next, we measured the latency of the response evoked in the brain with respect to the applied pulse. The Hampel method detected events in about 31% of trials, and the wavelet method in about 42% of trials. In addition, we found that movement had a statistical significant effect on the measured latency but only when L-dopa was administered to the animal. The differences in latency suggest that there may be a trajectory of neuronal activation in the brain in response of SCS. Our results may have implications in the design of more effective SCS strategies for the treatment of PD.

Status epilepticus is caused by a seizure lasting more than 5 minutes or several seizures in this time. For the detection of seizures, encephalograms are visually analyzed by doctors, but this has certain limitations, which can be reduced using algorithms that allow the identification of seizure patterns. Usually, the algorithms use all the channels of the electroencephalography, which causes more computational time. Therefore, the paper proposes an algorithm that seeks to verify that the use of fewer channels chosen for having less cross-correlation can lead to better seizure detection metrics. Of the classification algorithms used, XGBoost is the one that shows a more noticeable difference in sensitivity between 3 channels (80.64%) and 22 channels (78.19%). Also, ”FP1-F7”, ”A1-T3”, ”P3-O1” and ”FP1-F3” are the best channels for seizure detection. Research showed that using fewer channels selected by cross-correlation can improve seizure detection.

Oscillations are ubiquitous in the nervous system, from single neurons to whole brain networks. The link between neural oscillations and cognition and behaviour is actively investigated by cognitive and computational neuroscience. Biophysically motivated computational models, such as Wilson-Cowan [W-C], have contributed to the understanding of the dynamics of oscillatory neuronal networks. W-C describes mean field interactions between excitatory/inhibitory neural populations. Using Malkin’s Theorem we show the equivalence, under certain conditions, between the W-C model and the Kuramoto oscillators, with the advantage that the latter comprises fewer parameters. We construct a thirty-two nodes network of Kuramoto oscillators, coupled using two options: homogeneous (same strength in all connections) and heterogeneous (different values of coupling strengths). We characterized the Kuramoto network synchrony by measuring the Kuramoto order parameter, and the frequency spectrum of each oscillator using Welch’s periodograms. We characterized those two features as a function of number of nodes, their intrinsic frequency, and the global coupling parameter. Using variable intrinsic frequency between oscillators, we found that as we increase the number of nodes of the system, the global synchrony becomes dependent on the global coupling strength. Also, as global coupling increases, the frequency spectrum of each oscillator converges to the mean intrinsic frequency, similar to the case when the intrinsic frequency is equal for all nodes. We conclude that the Kuramoto order parameter alone is not enough of characterizing network dynamics, and that a distribution of intrinsic node frequency is important to generate the sort of network dynamics observed in brain imaging data.

Graph Signal Processing (GSP) is a framework for analyzing signals defined over a graph. Considering the electrodes used to record the electroencephalogram (EEG) as a sensor network makes it possible to use GSP to analyze EEG signals. Using the graph over which the signal is defined allows one to take advantage of a signal structure that is ignored by classic signal processing approaches. However, there are many details about how to use GSP to analyze the EEG that are not studied in the literature. Here we show an example of how to impute missing EEG data using GSP. We show that GSP allows reconstructing EEG missing data with a lower error than a classic approach based on radial basis functions, confirming that the underlying graph over a graph over which the signal is defined contains relevant information that can be exploited to improve a given signal processing task. By studying two approaches for building the graph (k-nearest neighbors and a thresholded Gaussian kernel) and the effect of its parameter, we highlight the importance of building the graph appropriately. These results show the potential of incorporating GPS techniques into the EEG processing pipeline.

Auscultation using stethoscopes allows the diagnosis of respiratory and cardiac diseases. However, these sounds interfere with each other both in time and frequency. In the case of recording heart sounds, it is possible to ask the patient to stop their breathing to perform auscultation and obtain a pure heart sound. But, in the case of lung sounds it is impossible to do the same. In this paper, a source separation method based on Non-negative Matrix Factorization (NMF) is used to decompose a signal into different components. The method proposed uses information from the estimated lung sound to reinsert the segments of interest into the original signal. The objective of this approximation is not to distort the segments of pure respiratory sound (free of heart sound). This method is compared to a base case of NMF decomposition on the raw signal. Three criteria for classifying the components based on the literature are also proposed, which will allow to indicate which component corresponds to each sound. The results were evaluated using temporal and spectral correlations, mean square error (MSE) and signal to distortion ratio (SDR) between the original respiratory signal and the respiratory signal estimated through the algorithm. It is shown that the best approximation is the NMF decomposition on the entire signal & replacing segments under different parameter variations.

Bayesian estimation based on an extended Kalman filter and a muscle-controlled biomechanical model of the vocal folds has demonstrated a viable, appealing method for obtaining physiologically relevant measures of glottal function (e.g., subglottal pressure, activation of laryngeal muscles, and vocal fold collision pressure) from synchronized recordings of calibrated high-speed videoendoscopy and oral airflow. In practice, the simultaneous recording of these signals is a cumbersome, uncomfortable procedure for patients. To simplify the experimental procedure while maintaining sufficient accuracy, we introduce a constrained Bayesian scheme that suitably incorporates physiological information about the subglottal pressure and uses the glottal area waveform as the sole observation. The proposed constrained extended Kalman filter produces reliable glottal airflow estimates compared to in-lab obtained waveforms, yielding root-mean-square errors lower than 63 mL/s, a performance similar to previous studies. These initial results provide advancement for the clinical assessment of vocal function.

Reverberant shear wave elastography is quantitative elastography approach that involves the generation of a reverberant field that propagates in all directions within the medium. Based on the generation of the field, the information about tissues stiffness can be obtained by measuring the shear wave speed (SWS). This approach has showed encouraging results for characterization of tissues such as the experiments already carried out in vivo. The aim of this study is to evaluate the quality of SWS estimators for acoustic reverberant fields in heterogeneous media to achieve this purpose, three estimators were tested: Curve fitting, Phase Gradient and a Convolutional Neural Network. Parameters such as time of processing, accuracy, resolution and contrast between background and inclusion were used to determine the best estimator. The results show the inclusion zone a maximum accuracy of 94.75% can be obtained with the deep learning estimator; while, in the background, the maximum accuracy was obtained with the phase gradient estimator and was 96.85%. Even though all three estimators show good results, the execution time, lateral resolution, contrast and CNR have better results with the deep learning estimator.

This work presents a watermarking algorithm applied to medical images by using the Steered Hermite Transform (SHT), the Singular Value Decomposition (SVD), and the Jigsaw transform (JS). The principal objective is to protect the patient’s information using imperceptible watermarking and preserve its diagnosis. Thus, the watermark imperceptibility is achieved using the high-order Steered Hermite coefficients, whereas the SVD decomposition and the JS ensure the watermark against attacks. We use the medicine symbol Caduceus as a watermark. The metrics employed to evaluate the algorithm’s performance are the Peak Signal-to-Noise Ratio (PSNR), the Mean Structural Similarity Index (MSSIM), and the Normalized Cross-Correlation (NCC). The evaluation metrics over the watermarked image show that it does not suffer quantitative and qualitative changes, and the extracted watermark was recovered successfully with high PSNR values. In addition, several watermark extraction tests were performed against geometric and common processing attacks. These tests show that the proposed algorithm is robust under critical conditions of attacks, for example, against nonlinear smoothing (median filter), high noise addition (Gaussian and Salt & Pepper noise), high compression rates (JPEG compression), rotation between 0 to 180 degree, and translations up to 100 pixels.

Brain structural changes in older adults over time can help identify and predict which individuals are at risk for neurodegenerative disorders and dementias. These trajectories are traditionally calculated by assessing localized rates of brain tissue atrophy from a longitudinal series of brain MRIs as compared to group averages. However, these methods do not preserve individual differences in brain structure, which may provide added information regarding risk. A map of how an individual’s brain may look at a given age - in the case of a normal, healthy, aging trajectory - may help to identify deviations and abnormalities when presented with a true scan at that age. Here, we consider estimating the age-related brain changes as a domain transfer problem. We develop a fully unsupervised generative adversarial network (GAN) with cycle consistency reconstruction losses, trained on cross-sectional brain MRI data from participants of the UK Biobank aged 45 to 81. We show that brain MRIs for males and females at a given age can be predicted by converting the content information encoded in a T1-weighted MRI (i.e., the individual’s identifying anatomical features), accompanied by adding the style (age/sex) information from a reference group. Results on the PREVENT-AD cohort demonstrated that our style-encoding domain transfer model can predict follow-up brain MRIs, successfully, without relying on longitudinal data from the subjects. We show how deviations from the predicted images are indicative of factors related to neurodegenerative disease risk.

Nowadays eye diseases that are not treated in a timely manner can lead to blindness in the patient. Diabetic retinopathy and retinopathy of prematurity are a couple of conditions considered to be the main causes of blindness in both adults and children. The technique used to date to verify the status of the retina is a qualitative analysis by an ophthalmological expert of fundus images. However, this is entirely based on the experience acquired by the physician and being able to detect changes in the vascular structure of the retina is a great challenge which can be addressed through technology. This paper presents a novel method to carry out the numerical modeling of the major temporal arcade using orthogonal polynomials of Legendre, Chebyshev and Laguerre through a genetic algorithm that helps to determine the coefficients of the linear combination of each one. A set of twenty fundus images already outlined by an expert was used, which were processed by the algorithm, generating an adjustment curve on the set of pixels of the Major Temporal Arcade. The results obtained were compared with three existing methodologies in the literature by using two metrics, emerging the Legendre polynomials as the most suitable for modeling, as a consequence of the low values obtained in the metrics compared to the other methods.

Diabetic retinopathy (DR) is one of the leading causes of blindness in the working-age population of developed countries, caused by a side effect of diabetes that reduces the blood supply to the retina. Deep neural networks have been widely used in automated systems for DR classification on eye fundus images. However, these models need a large number of annotated images. In the medical domain, annotations from experts are costly, tedious, and time-consuming; as a result, a limited number of annotated images are available. This paper presents a semi-supervised method that leverages unlabeled images and labeled ones to train a model that detects diabetic retinopathy. The proposed method uses unsupervised pretraining via self-supervised learning followed by supervised fine-tuning with a small set of labeled images and knowledge distillation to increase the performance in a classification task. This method was evaluated on the EyePACS test and Messidor-2 dataset achieving 0.94 and 0.89 AUC respectively using only 2% of EyePACS train labeled images.

In early life, the neurocranium undergoes rapid changes to accommodate the expanding brain. Neurocranial maturation can be disrupted by developmental abnormalities and environmental factors such as sleep position. To establish a baseline for the early detection of anomalies, it is important to understand how this structure typically grows in healthy children. Here, we designed a deep neural network pipeline NEC-NET, including segmentation and classification, to analyze the normative development of the neurocranium in T1 MR images from healthy children aged 12 to 60 months old. The pipeline optimizes the segmentation of the neurocranium and shows the preliminary results of age-based regional differences among infants.

Deep neural networks show great promise for classifying brain diseases and making prognostic assessments based on neuroimaging data, but large, labeled training datasets are often required to achieve high predictive accuracy. Here we evaluated a range of transfer learning or pre-training strategies to create useful MRI representations for downstream tasks that lack large amounts of training data, such as Alzheimer’s disease (AD) classification. To test our proposed pretraining strategies, we analyzed 4,098 3D T1-weighted brain MRI scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort and independently validated with an out-of-distribution test set of 600 scans from the Open Access Series of Imaging Studies (OASIS3) cohort for detecting AD. First, we trained 3D and 2D convolutional neural network (CNN) architectures. We tested combinations of multiple pre-training strategies based on (1) supervised, (2) contrastive learning, and (3) self-supervised learning - using pre-training data within versus outside the MRI domain. In our experiments, the 3D CNN pre-trained with contrastive learning provided the best overall results - when fine-tuned on T1-weighted scans for AD classification - outperformed the baseline by 2.8% when trained with all of the training data from ADNI. We also show test performance as a function of the training dataset size and the chosen pre-training method. Transfer learning offered significant benefits in low data regimes, with a performance boost of 7.7%. When the pretrained model was used for AD classification, we were able to visualize an improved clustering of test subjects' diagnostic groups, as illustrated via a uniform manifold approximation (UMAP) projection of the high-dimensional model embedding space. Further, saliency maps indicate the additional activation regions in the brain scan using pretraining, that then maximally contributed towards the final prediction score.

The objective of the present work is the implementation of a Deep Learning approach using convolutional neural networks that automatically detects the presence of crackles or wheezes, anomalous respiratory sounds present during the inspiration and expiration. In addition, a proposal of a capture mechanism composed by an electret microphone and an acoustic coupler is presented. All of this, with the purpose of devising a support tool in the early respiratory disease diagnosis. Among the main conclusions, it was found that the most appropriate indicator for the model evaluation was sensitivity, where a value of 94,17% was obtained for the validation set, which shows an adequate performance. Additionally, the highest number of erroneous classifications occurred with the crackles, while the lowest in the wheezes, concluding that the system is more effective detecting the second kind of sound. In perspective, the development of a classification algorithm is proposed by taking advantage of frequential and temporal analysis, it manages to find the frequency range and respiratory cycle stage where the anomalous sound happened with the purpose of reaching more specific and precise diagnoses.