This paper presents some preliminary results using Landsat and Worldview images for change detection. The studied area had some significant changes such as construction of buildings between May 2014 and October 2015. We investigated several simple, practical, and effective approaches to change detection. For Landsat images, we first performed pansharpening to enhance the resolution to 15 meters. We then performed a chronochrome covariance equalization between two images. The residual between the two equalized images was then analyzed using several simple algorithms such as direct subtraction and global Reed-Xiaoli (GRX) detector. Experimental results using actual Landsat images clearly demonstrated that the proposed methods are effective. For Worldview images, we used pansharpened images with only four bands for change detection. The performance of the aforementioned algorithms is comparable to that of a commercial package developed by Digital Globe.
Pansharpened Landsat images have 15 m spatial resolution with 16-day revisit periods. On the other hand, Worldview images have 0.5 m resolution after pansharpening but the revisit times are uncertain. We present some preliminary results for a challenging image fusion problem that fuses Landsat and Worldview (WV) images to yield a high temporal resolution image sequence at the same spatial resolution of WV images. Since the spatial resolution between Landsat and Worldview is 30 to 1, our preliminary results are mixed in that the objective performance metrics such as peak signal-to-noise ratio (PSNR), correlation coefficient (CC), etc. sometimes showed good fusion performance, but at other times showed poor results. This indicates that more fusion research is still needed in this niche application.
With the rapid development of earth observation technology, remote sensing images have played more important roles, because the high resolution images can provide the original data for object recognition, disaster investigation, and so on. When a disastrous earthquake breaks out, a large number of roads could be damaged instantly. There are a lot of approaches about road extraction, such as region growing, gray threshold, and k-means clustering algorithm. We could not obtain the undamaged roads with these approaches, if the trees or their shadows along the roads are difficult to be distinguished from the damaged road. In the paper, a method is presented to extract the damaged road with high resolution aerial image of post-earthquake. Our job is to extract the damaged road and the undamaged with the aerial image. We utilized the mathematical morphology approach and the k-means clustering algorithm to extract the road. Our method was composed of four ingredients. Firstly, the mathematical morphology filter operators were employed to remove the interferences from the trees or their shadows. Secondly, the k-means algorithm was employed to derive the damaged segments. Thirdly, the mathematical morphology approach was used to extract the undamaged road; Finally, we could derive the damaged segments by overlaying the road networks of pre-earthquake. Our results showed that the earthquake, broken in Yaan, was disastrous for the road, Therefore, we could take more measures to keep it clear.
Ruoergai Conservation area belongs to the Aba Tibetan and Qiang Autonomous Prefecture, Sichuan Province. In recent decades, the wetland and grassland have degraded very seriously. In this paper, based on GIS spatial analysis techniques, we utilized the Landsat 7 ETM + images in 2007, 2009, and 2011, to extract the land use/land cover change in Ruoergai. Firstly, the images were enhanced, mosaicked, and subset. Secondly, the supervised and unsupervised classification method were used to derive the land use/land cover change information in Ruoergai. Finally, the land use / land cover change in Ruoergai was analyzed between 2007 and 2011 .
Our results were listed below:
(1)The area of water body and swamp in Ruoergai reduced by 52.346% between 2007 and 2011.
(2)The area of grass land in Ruoergai decreased by 12.754% between 2007 and 2011.
(3) The area of woodland in Ruoergai reduced by 3.224% between 2007 and 2011.
(4) The area of bare land, cultivated land and construction land increased by 6.647% between 2007 and 2011.
We present a sparse coding based dense feature representation model (a preliminary version of the paper was presented at the SPIE Remote Sensing Conference, Dresden, Germany, 2013) for hyperspectral image (HSI) classification. The proposed method learns a new representation for each pixel in HSI through the following four steps: sub-band construction, dictionary learning, encoding, and feature selection. The new representation usually has a very high dimensionality requiring a large amount of computational resources. We applied the l1/lq regularized multiclass logistic regression technique to reduce the size of the new representation. We integrated the method with a linear support vector machine (SVM) and a composite kernels SVM (CKSVM) to discriminate different types of land cover. We evaluated the proposed algorithm on three well-known HSI datasets and compared our method to four recently developed classification methods: SVM, CKSVM, simultaneous orthogonal matching pursuit, and image fusion and recursive filtering. Experimental results show that the proposed method can achieve better overall and average classification accuracies with a much more compact representation leading to more efficient sparse models for HSI classification.
MR diffusion tensor imaging (DTI) technique together with traditional T1 or T2 weighted MRI scans supplies rich information sources for brain cancer diagnoses. These images form large-scale, high-dimensional data sets. Due to the fact that significant correlations exist among these images, we assume low-dimensional geometry data structures (manifolds) are embedded in the high-dimensional space. Those manifolds might be hidden from radiologists because it is challenging for human experts to interpret high-dimensional data. Identification of the manifold is a critical step for successfully analyzing multimodal MR images.
We have developed various manifold learning algorithms (Tran et al. 2011; Tran et al. 2013) for medical image analysis. This paper presents a comparative study of an incremental manifold learning scheme (Tran. et al. 2013) versus the deep learning model (Hinton et al. 2006) in the application of brain tumor progression prediction. The incremental manifold learning is a variant of manifold learning algorithm to handle large-scale datasets in which a representative subset of original data is sampled first to construct a manifold skeleton and remaining data points are then inserted into the skeleton by following their local geometry. The incremental manifold learning algorithm aims at mitigating the computational burden associated with traditional manifold learning methods for large-scale datasets. Deep learning is a recently developed multilayer perceptron model that has achieved start-of-the-art performances in many applications. A recent technique named "Dropout" can further boost the deep model by preventing weight coadaptation to avoid over-fitting (Hinton et al. 2012).
We applied the two models on multiple MRI scans from four brain tumor patients to predict tumor progression and compared the performances of the two models in terms of average prediction accuracy, sensitivity, specificity and precision. The quantitative performance metrics were calculated as average over the four patients. Experimental results show that both the manifold learning and deep neural network models produced better results compared to using raw data and principle component analysis (PCA), and the deep learning model is a better method than manifold learning on this data set. The averaged sensitivity and specificity by deep learning are comparable with these by the manifold learning approach while its precision is considerably higher. This means that the predicted abnormal points by deep learning are more likely to correspond to the actual progression region.
Isomap is a classical manifold learning approach that preserves geodesic distance of nonlinear data sets. One of the main drawbacks of this method is that it is susceptible to leaking, where a shortcut appears between normally separated portions of a manifold. We propose an adaptive graph construction approach that is based upon the sparsity property of the ℓ1 norm. The ℓ1 enhanced graph construction method replaces k-nearest neighbors in the classical approach. The proposed algorithm is first tested on the data sets from the UCI data base repository which showed that the proposed approach performs better than the classical approach. Next, the proposed approach is applied to two image data sets and achieved improved performances over standard Isomap.
Vehicle detection is a very important task for intelligent transportation system. In this paper, a method with mathematical morphology and template matching is presented to detect the crowded vehicles of parking lot with high resolution aerial image. Our experimental results with high resolution aerial image showed that the graded image, with the spatial resolution of 1×1ft, could greatly reduce the calculation time, but with the same accuracy as the original image with the spatial resolution of 0.5×0.5ft .
We present a two-step algorithm for the detection of seafloor propeller seagrass scars in shallow water using panchromatic images. The first step is to classify image pixels into scar and non-scar categories based on a sparse coding algorithm. The first step produces an initial scar map in which false positive scar pixels may be present. In the second step, local orientation of each detected scar pixel is computed using the morphological directional profile, which is defined as outputs of a directional filter with a varying orientation parameter. The profile is then utilized to eliminate false positives and generate the final scar detection map. We applied the algorithm to a panchromatic image captured at the Deckle Beach, Florida using the WorldView2 orbiting satellite. Our results show that the proposed method can achieve <90% accuracy on the detection of seagrass scars.
We present a sparse coding based spectral-spatial classification model for hyperspectral image (HSI) datasets. The
proposed method consists of an efficient sparse coding method in which the l1/lqregularized multi-class logistic regression technique was utilized to achieve a compact representation of hyperspectral image pixels for land cover
classification. We applied the proposed algorithm to a HSI dataset collected at the Kennedy Space Center and compared
our algorithm to a recently proposed method, Gaussian process maximum likelihood (GP-ML) classifier. Experimental
results show that the proposed method can achieve significantly better performances than the GP-ML classifier when
training data is limited with a compact pixel representation, leading to more efficient HSI classification systems.
Improvement in sensor technology such as charge-coupled devices (CCD) as well as constant incremental improvements
in storage space has enabled the recording and storage of video more prevalent and lower cost than ever before.
However, the improvements in the ability to capture and store a wide array of video have required additional manpower
to translate these raw data sources into useful information. We propose an algorithm for automatically detecting
anomalous movement patterns within full motion video thus reducing the amount of human intervention required to
make use of these new data sources. The proposed algorithm tracks all of the objects within a video sequence and
attempts to cluster each object's trajectory into a database of existing trajectories. Objects are tracked by first
differentiating them from a Gaussian background model and then tracked over subsequent frames based on a
combination of size and color. Once an object is tracked over several frames, its trajectory is calculated and compared
with other trajectories earlier in the video sequence. Anomalous trajectories are differentiated by their failure to cluster
with other well-known movement patterns. Adding the proposed algorithm to an existing surveillance system could
increase the likelihood of identifying an anomaly and allow for more efficient collection of intelligence data.
Additionally, by operating in real-time, our algorithm allows for the reallocation of sensing equipment to those areas
most likely to contain movement that is valuable for situational awareness.
Many applications require to register images within subpixel accuracy like computer vision especially super-resolution
(SR) where the estimated subpixel shifts are very crucial in the reconstruction and restoration of SR
images. In our work we have an optical sensor that is mounted on an unmanned airborne vehicle (UAV) and
captures a set of images that contain sufficient overlapped area required to reconstruct a SR image. Due to the
wind, The UAV may encounter rotational effects such as yaw, pitch and roll which can distort the acquired as
well as processed images with shear, tilt or perspective distortions. In this paper we propose a hybrid algorithm
to register these UAV images within subpixel accuracy to feed them in a SR reconstruction step. Our algorithm
consists of two steps. The first step uses scale invariant feature transform (SIFT) to correct the distorted images.
Because the resultant images are not registered to a subpixel precision, the second step registers the images
using a fast Fourier transform (FFT) based method that is both efficient and robust to moderate noise and lens
optical blur. Our FFT based method reduces the dimensionality of the Fourier matrix of the cross correlation
and uses a forward and backward search in order to obtain an accurate estimation of the subpixel shifts. We
discuss the relation between the dimensionality reduction factors and the image shifts as well as propose criteria
that can be used to optimally select these factors. Finally, we compare the results of our approach to other
subpixel techniques in terms of their efficiency and computational speed.
Many techniques have been recently developed for classification of hyperspectral images (HSI) including support vector
machines (SVMs), neural networks and graph-based methods. To achieve good performances for the classification, a
good feature representation of the HSI is essential. A great deal of feature extraction algorithms have been developed
such as principal component analysis (PCA) and independent component analysis (ICA). Sparse coding has recently
shown state-of-the-art performances in many applications including image classification. In this paper, we present a
feature extraction method for HSI data motivated by a recently developed sparse coding based image representation
technique. Sparse coding consists of a dictionary learning step and an encoding step. In the learning step, we compared
two different methods, L1-penalized sparse coding and random selection for the dictionary learning. In the encoding step,
we utilized a soft threshold activation function to obtain feature representations for HSI. We applied the proposed
algorithm to a HSI dataset collected at the Kennedy Space Center (KSC) and compared our results with those obtained
by a recently proposed method, supervised locally linear embedding weighted k-nearest-neighbor (SLLE-WkNN)
classifier. We have achieved better performances on this dataset in terms of the overall accuracy with a random
dictionary. We conclude that this simple feature extraction framework might lead to more efficient HSI classification
Super-resolution (SR) refers to reconstructing a single high resolution (HR) image from a set of subsampled,
blurred and noisy low resolution (LR) images. The reconstructed image suffers from degradations such as
blur, aliasing, photo-detector noise and registration and fusion error. Wiener filter can be used to remove
artifacts and enhance the visual quality of the reconstructed images. In this paper, we introduce a new fast
stochasticWiener filter for SR reconstruction and restoration that can be implemented efficiently in the frequency
domain. Our derivation depends on the continuous-discrete-continuous (CDC) model that represents most of
the degradations encountered during the image-gathering and image-display processes. We incorporate a new
parameter that accounts for LR images registration and fusion errors. Also, we speeded up the performance
of the filter by constraining it to work on small patches of the images. Beside this, we introduce two figures
of merits: information rate and maximum realizable fidelity, which can be used to assess the visual quality
of the resultant images. Simulations and experimental results demonstrate that the derived Wiener filter that
can be implemented efficiently in the frequency domain can reduce aliasing, blurring, and noise and result in a
sharper reconstructed image. Also, Quantitative assessment using the proposed figures coincides with the visual
qualitative assessment. Finally, we evaluate our filter against other SR techniques and its results were very
In super-resolution (SR), a set of degraded low-resolution (LR) images are used to reconstruct a higher-resolution image that suffers from acquisition degradations. One way to boost SR images visual quality is to use restoration filters to remove reconstructed images artifacts. We propose an efficient method to optimally allocate the LR pixels on the high-resolution grid and introduce a mathematical derivation of a stochastic Wiener filter. It relies on the continuous-discrete-continuous model and is constrained by the periodic and nonperiodic interrelationships between the different frequency components of the proposed SR system. We analyze an end-to-end model and formulate the Wiener filter as a function of the parameters associated with the proposed SR system such as image gathering and display response indices, system average signal-to-noise ratio, and inter-subpixel shifts between the LR images. Simulation and experimental results demonstrate that the derived Wiener filter with the optimal allocation of LR images results in sharper reconstruction. When compared with other SR techniques, our approach outperforms them in both quality and computational time.
Efficient application of wound treatment procedures is vital in both emergency room and battle zone scenes. In order to
train first responders for such situations, physical casualty simulation kits, which are composed of tens of individual
items, are commonly used. Similar to any other training scenarios, computer simulations can be effective means for
wound treatment training purposes. For immersive and high fidelity virtual reality applications, realistic 3D models are
key components. However, creation of such models is a labor intensive process. In this paper, we propose a procedural
wound geometry generation technique that parameterizes key simulation inputs to establish the variability of the training
scenarios without the need of labor intensive remodeling of the 3D geometry. The procedural techniques described in
this work are entirely handled by the graphics processing unit (GPU) to enable interactive real-time operation of the
simulation and to relieve the CPU for other computational tasks. The visible human dataset is processed and used as a
volumetric texture for the internal visualization of the wound geometry. To further enhance the fidelity of the simulation,
we also employ a surface flow model for blood visualization. This model is realized as a dynamic texture that is
composed of a height field and a normal map and animated at each simulation step on the GPU. The procedural wound
geometry and the blood flow model are applied to a thigh model and the efficiency of the technique is demonstrated in a
virtual surgery scene.
Proc. SPIE. 8135, Applications of Digital Image Processing XXXIV
KEYWORDS: Principal component analysis, 3D image reconstruction, Data modeling, Cameras, Databases, Image processing, 3D modeling, Light sources and illumination, Reconstruction algorithms, 3D image processing
3D face modeling has been one of the greatest challenges for researchers in computer graphics for many years. Various
methods have been used to model the shape and texture of faces under varying illumination and pose conditions from a
single given image. In this paper, we propose a novel method for the 3D face synthesis and reconstruction by using a
simple and efficient global optimizer. A 3D-2D matching algorithm which employs the integration of the 3D morphable
model (3DMM) and the differential evolution (DE) algorithm is addressed. In 3DMM, the estimation process of fitting
shape and texture information into 2D images is considered as the problem of searching for the global minimum in a
high dimensional feature space, in which optimization is apt to have local convergence. Unlike the traditional scheme
used in 3DMM, DE appears to be robust against stagnation in local minima and sensitiveness to initial values in face
reconstruction. Benefitting from DE's successful performance, 3D face models can be created based on a single 2D
image with respect to various illuminating and pose contexts. Preliminary results demonstrate that we are able to
automatically create a virtual 3D face from a single 2D image with high performance. The validation process shows that
there is only an insignificant difference between the input image and the 2D face image projected by the 3D model.
Super-resolution (SR) is the process of obtaining a higher resolution image from a set of lower resolution (LR)
blurred and noisy images. One may, then, envision a scenario where a set of LR images is acquired with a
sensor on a moving platform. In such a case, an SR image can be reconstructed in an area of sufficient overlap
between the LR images which generally have a relative shift with respect to each other by subpixel amounts.
The visual quality of the SR image is affected by many factors such as the optics blur, the inherent signalto-
noise ratio of the system, quantization artifacts, the number of scenels (scene elements) i.e., the number of
overlapped images used for SR reconstruction within the SR grid and their relative arrangement. In most cases
of microscanning, the subpixel shifts between the LR images are pre-determined: hence the number of the scenels
within the SR grid and their relative positions with respect to each other are known and, as a result, can be used
in obtaining the reconstructed SR image with high quality. However, the LR images may have relative shifts
that are unknown. This random pattern of subpixel shifts can lead to unpleasant visual quality, especially at
the edges of the reconstructed SR image. Also, depending on the available number of the LR images and their
relative positions, it may be possible to produce SR only along a single dimension diagonal, horizontal or vertical
and use interpolation in the orthogonal dimension because there isn't sufficient information to produce a full 2D
image. We investigate the impact of the number of overlapped regions and their relative arrangement on the
quality of the SR images, and propose a technique that optimally allocates the available LR scenels to the SR
grid in order to minimize the expected unpleasant visual artifacts.
Pose and illumination are identified as major problems in 2D face recognition (FR). It has been theoretically proven that
the more diversified instances in the training phase, the more accurate and adaptable the FR system appears to be. Based
on this common awareness, researchers have developed a large number of photographic face databases to meet the
demand for data training purposes. In this paper, we propose a novel scheme for 2D face database diversification based
on 3D face modeling and computer graphics techniques, which supplies augmented variances of pose and illumination.
Based on the existing samples from identical individuals of the database, a synthesized 3D face model is employed to
create composited 2D scenarios with extra light and pose variations. The new model is based on a 3D Morphable Model
(3DMM) and genetic type of optimization algorithm. The experimental results show that the complemented instances
obviously increase diversification of the existing database.
Proc. SPIE. 8056, Visual Information Processing XX
KEYWORDS: Visual process modeling, Detection and tracking algorithms, Visualization, Imaging systems, Cameras, Sensors, Image enhancement, Human vision and color perception, CAD systems, Solid modeling
An automatic landing site detection algorithm is proposed for aircraft emergency landing. Emergency landing
is an unplanned event in response to emergency situations. If, as is unfortunately usually the case, there is
no airstrip or airfield that can be reached by the un-powered aircraft, a crash landing or ditching has to be
carried out. Identifying a safe landing site is critical to the survival of passengers and crew. Conventionally,
the pilot chooses the landing site visually by looking at the terrain through the cockpit. The success of this
vital decision greatly depends on the external environmental factors that can impair human vision, and on
the pilot's flight experience that can vary significantly among pilots. Therefore, we propose a robust, reliable
and efficient algorithm that is expected to alleviate the negative impact of these factors. We present only the
detection mechanism of the proposed algorithm and assume that the image enhancement for increased visibility,
and image stitching for a larger field-of-view have already been performed on the images acquired by aircraftmounted
cameras. Specifically, we describe an elastic bound detection method which is designed to position
the horizon. The terrain image is divided into non-overlapping blocks which are then clustered according to a
"roughness" measure. Adjacent smooth blocks are merged to form potential landing sites whose dimensions are
measured with principal component analysis and geometric transformations. If the dimensions of the candidate
region exceed the minimum requirement for safe landing, the potential landing site is considered a safe candidate
and highlighted on the human machine interface. At the end, the pilot makes the final decision by confirming
one of the candidates, also considering other factors such as wind speed and wind direction, etc. Preliminary
results show the feasibility of the proposed algorithm.
At various stage of progression, most brain tumors are not homogenous. In this presentation, we retrospectively studied
the distribution of ADC values inside tumor volume during the course of tumor treatment and progression for a selective
group of patients who underwent an anti-VEGF trial. Complete MRI studies were obtained for this selected group of
patients including pre- and multiple follow-up, post-treatment imaging studies. In each MRI imaging study, multiple
scan series were obtained as a standard protocol which includes T1, T2, T1-post contrast, FLAIR and DTI derived
images (ADC, FA etc.) for each visit. All scan series (T1, T2, FLAIR, post-contrast T1) were registered to the
corresponding DTI scan at patient's first visit. Conventionally, hyper-intensity regions on T1-post contrast images are
believed to represent the core tumor region while regions highlighted by FLAIR may overestimate tumor size. Thus we
annotated tumor regions on the T1-post contrast scans and ADC intensity values for pixels were extracted inside tumor
regions as defined on T1-post scans. We fit a mixture Gaussian (MG) model for the extracted pixels using the
Expectation-Maximization (EM) algorithm, which produced a set of parameters (mean, various and mixture coefficients)
for the MG model. This procedure was performed for each visits resulting in a series of GM parameters. We studied the
parameters fitted for ADC and see if they can be used as indicators for tumor progression. Additionally, we studied the
ADC characteristics in the peri-tumoral region as identified by hyper-intensity on FLAIR scans. The results show that
ADC histogram analysis of the tumor region supports the two compartment model that suggests the low ADC value subregion
corresponding to densely packed cancer cell while the higher ADC value region corresponding to a mixture of
viable and necrotic cells with superimposed edema. Careful studies of the composition and relative volume of the two
compartments in tumor region may provide some insights in the early assessment of tumor response to therapy for
recurrence brain cancer patients.
In a recent study , we investigated the feasibility of predicting brain tumor progression based on multiple MRI series
and we tested our methods on seven patients' MRI images scanned at three consecutive visits A, B and C. Experimental
results showed that it is feasible to predict tumor progression from visit A to visit C using a model trained by the
information from visit A to visit B. However, the trained model failed when we tried to predict tumor progression from
visit B to visit C, though it is clinically more important. Upon a closer look at the MRI scans revealed that histograms of
MRI scans such as T1, T2, FLAIR etc taken at different times have slight shifts or different shapes. This is because those
MRI scans are qualitative instead of quantitative so MRI scans taken at different times or by different scanners might
have slightly different scales or have different homogeneities in the scanning region. In this paper, we proposed a
method to overcome this difficulty. The overall goal of this study is to assess brain tumor progression by exploring seven
patients' complete MRI records scanned during their visits in the past two years. There are ten MRI series in each visit,
including FLAIR, T1-weighted, post-contrast T1-weighted, T2-weighted and five DTI derived MRI volumes: ADC, FA,
Max, Min and Middle Eigen Values. After registering all series to the corresponding DTI scan at the first visit, we
applied a histogram matching algorithm to non-DTI MRI scans to match their histograms to those of the corresponding
MRI scans at the first visit. DTI derived series are quantitative and do not require the histogram matching procedure. A
machine learning algorithm was then trained using the data containing information from visit A to visit B, and the
trained model was used to predict tumor progression from visit B to visit C. An average of 72% pixel-wise accuracy was
achieved for tumor progression prediction from visit B to visit C.
In this paper, we present a texture analysis based method for diagnosing the Basal Cell Carcinoma (BCC) skin cancer
using optical images taken from the suspicious skin regions. We first extracted the Run Length Matrix and Haralick
texture features from the images and used a feature selection algorithm to identify the most effective feature set for the
diagnosis. We then utilized a Multi-Layer Perceptron (MLP) classifier to classify the images to BCC or normal cases.
Experiments showed that detecting BCC cancer based on optical images is feasible. The best sensitivity and specificity
we achieved on our data set were 94% and 95%, respectively.
For the early detection of prostate cancer, the analysis of the Prostate-specific antigen (PSA) in serum is currently the
most popular approach. However, previous studies show that 15% of men have prostate cancer even their PSA
concentrations are low. MALDI Mass Spectrometry (MS) proves to be a better technology to discover molecular tools
for early cancer detection. The molecular tools or peptides are termed as biomarkers. Using MALDI MS data from
prostate tissue samples, prostate cancer biomarkers can be identified by searching for molecular or molecular
combination that can differentiate cancer tissue regions from normal ones. Cancer tissue regions are usually identified by
pathologists after examining H&E stained histological microscopy images. Unfortunately, histopathological examination
is currently done on an adjacent slice because the H&E staining process will change tissue's protein structure and it will
derogate MALDI analysis if the same tissue is used, while the MALDI imaging process will destroy the tissue slice so
that it is no longer available for histopathological exam. For this reason, only the most confident cancer region resulting
from the histopathological examination on an adjacent slice will be used to guide the biomarker identification. It is
obvious that a better cancer boundary delimitation on the MALDI imaging slice would be beneficial. In this paper, we
proposed methods to predict the true cancer boundary, using the MALDI MS data, from the most confident cancer region
given by pathologists on an adjacent slice.
A machine learning technique is presented for assessing brain tumor progression by exploring six patients' complete
MRI records scanned during their visits in the past two years. There are ten MRI series, including diffusion tensor image
(DTI), for each visit. After registering all series to the corresponding DTI scan at the first visit, annotated normal and
tumor regions were overlaid. Intensity value of each pixel inside the annotated regions were then extracted across all of
the ten MRI series to compose a 10 dimensional vector. Each feature vector falls into one of three categories:normal,
tumor, and normal but progressed to tumor at a later time. In this preliminary study, we focused on the trend of brain
tumor progression during three consecutive visits, i.e., visit A, B, and C. A machine learning algorithm was trained using
the data containing information from visit A to visit B, and the trained model was used to predict tumor progression from
visit A to visit C. Preliminary results showed that prediction for brain tumor progression is feasible. An average of
80.9% pixel-wise accuracy was achieved for tumor progression prediction at visit C.
Prostate cancer is the second most common type of cancer among men in US . Traditionally, prostate cancer
diagnosis is made by the analysis of prostate-specific antigen (PSA) levels and histopathological images of biopsy
samples under microscopes. Proteomic biomarkers can improve upon these methods. MALDI molecular spectra imaging
is used to visualize protein/peptide concentrations across biopsy samples to search for biomarker candidates.
Unfortunately, traditional processing methods require histopathological examination on one slice of a biopsy sample
while the adjacent slice is subjected to the tissue destroying desorption and ionization processes of MALDI. The highest
confidence tumor regions gained from the histopathological analysis are then mapped to the MALDI spectra data to
estimate the regions for biomarker identification from the MALDI imaging. This paper describes a process to provide a
significantly better estimate of the cancer tumor to be mapped onto the MALDI imaging spectra coordinates using the
high confidence region to predict the true area of the tumor on the adjacent MALDI imaged slice.
Clinical MRI images are generally corrupted by random noise during acquisition with blurred subtle structure features.
Many denoising methods have been proposed to remove noise from corrupted images at the expense of distorted
structure features. Therefore, there is always compromise between removing noise and preserving structure information
for denoising methods. For a specific denoising method, it is crucial to tune it so that the best tradeoff can be obtained. In
this paper, we define several cost functions to assess the quality of noise removal and that of structure information
preserved in the denoised image. Strength Pareto Evolutionary Algorithm 2 (SPEA2) is utilized to simultaneously
optimize the cost functions by modifying parameters associated with the denoising methods. The effectiveness of the
algorithm is demonstrated by applying the proposed optimization procedure to enhance the image denoising results using
block matching and 3D collaborative filtering. Experimental results show that the proposed optimization algorithm can
significantly improve the performance of image denoising methods in terms of noise removal and structure information preservation.
Prostate cancer is the most common type of cancer and the second leading cause of cancer death among men in US1.
Quantitative assessment of prostate histology provides potential automatic classification of prostate lesions and
prediction of response to therapy. Traditionally, prostate cancer diagnosis is made by the analysis of prostate-specific
antigen (PSA) levels and histopathological images of biopsy samples under microscopes. In this application, we utilize a
texture analysis method based on the run-length matrix for identifying tissue abnormalities in prostate histology. A tissue
sample was collected from a radical prostatectomy, H&E fixed, and assessed by a pathologist as normal tissue or
prostatic carcinoma (PCa). The sample was then subsequently digitized at 50X magnification. We divided the digitized
image into sub-regions of 20 X 20 pixels and classified each sub-region as normal or PCa by a texture analysis method.
In the texture analysis, we computed texture features for each of the sub-regions based on the Gray-level Run-length
Matrix(GL-RLM). Those features include LGRE, HGRE and RPC from the run-length matrix, mean and standard
deviation of the pixel intensity. We utilized a feature selection algorithm to select a set of effective features and used a
multi-layer perceptron (MLP) classifier to distinguish normal from PCa. In total, the whole histological image was
divided into 42 PCa and 6280 normal regions. Three-fold cross validation results show that the proposed method
achieves an average classification accuracy of 89.5% with a sensitivity and specificity of 90.48% and 89.49%,
Purpose: Diffusion tensor imaging (DTI) is an inherently quantitative imaging technique that measures the
diffusivities of water molecules in tissue. However, the accuracy of DTI measurements depends on many
factors such S/N ratio and magnet field strength. Therefore, before quantitative assessment of tumor
progression based on DTI metric changes can be made with confidence, one have to assess the accuracy or
variance in the DTI metrics. This is especially important for multi-institutional clinical trials or for large
institutions where patients may be imaged on multiple MR scanners at multiple times in follow up studies.
In this presentation, we studied the feasibility of using CSF as an internal QC marker for data acquisition
and processing qualities. Method: ADC and FA of CSF for brain tumor patients' DTI studies (total of 85
scans over three years) were analyzed. In addition, a phantom was used to check the inherent variations of
the MR systems. Results: The results show that the coefficient of variations for ADC and FA are 8.4% and
13.2% in CSF among all patients. For all DTI scans done on 1.5 T scanners, they are 7.4% and 9.1%, while
for 3T they are 9.8% and 18% respectively. Conclusion: CSF can be used as an internal QC measure of the
DTI acquisition accuracy and consistency among longitude studies on patients, making it a potentially
useful in multi-institutional trials.
The height and width of colonic polyps are important characteristics to evaluate the status and malignancy of polyps. We borrow the idea from geographic information systems to employ topographic height maps to compute the polyp height and width. The height map is generated using a ray-casting algorithm through an orthogonal projection. A concentric index is devised to gauge the quality of the height map and is maximized in a multi-scale spiral spherical search for the
optimal projection. We then locate the polyp tip and neck using directional height profiles, and derive height and width
measurement based on geometrical analysis. We manually measured the height and width of 58 polyps and performed paired t-tests between manual measurement and height map measurement. The test shows that Pearson correlation is 0.742 and P(T<=t) is 0.01 for height measurement; and Pearson correlation is 0.663 and P(T<=t) is 0.002 for width measurement.
We evaluated a Pareto front-based multi-objective evolutionary algorithm for optimizing our CT colonography
(CTC) computer-aided detection (CAD) system. The system identifies colonic polyps based on curvature and
volumetric based features, where a set of thresholds for these features was optimized by the evolutionary algorithm.
We utilized a two-fold cross-validation (CV) method to test if the optimized thresholds can be generalized
to new data sets. We performed the CV method on 133 patients; each patient had a prone and a supine scan.
There were 103 colonoscopically confirmed polyps resulting in 188 positive detections in CTC reading from either
the prone or the supine scan or both. In the two-fold CV, we randomly divided the 133 patients into two
cohorts. Each cohort was used to obtain the Pareto front by a multi-objective genetic algorithm, where a set of
optimized thresholds was applied on the test cohort to get test results. This process was repeated twice so that
each cohort was used in the training and testing process once. We averaged the two training Pareto fronts as
our final training Pareto front and averaged the test results from the two runs in the CV as our final test results.
Our experiments demonstrated that the averaged testing results were close to the mean Pareto front determined
from the training process. We conclude that the Pareto front-based algorithm appears to be generalizable to
new test data.
Multiple classifiers working collaboratively can usually achieve better performance than any single classifier working
independently. Our CT colonography computer-aided detection (CAD) system uses support vector machines (SVM) as
the classifier. In this paper, we developed and evaluated two schemes to collaboratively apply multiple SVMs in the
same CAD system. One is to put the classifiers in a sequence (SVM sequence) and apply them one after another; the
other is to put the classifiers in a committee (SVM committee) and use the committee decision for the classification. We
compared the sequence order (best-first, worst-first and random) in the SVM sequence and two decision functions in the
SVM committee (majority vote and sum probability). The experiments were conducted on 786 CTC datasets, with 63
polyp detections. We used 10-fold cross validation to generate the FROC curves, and conducted 100 bootstraps to
evaluate the performance variation. The result showed that collaborative classifiers performed much better than
individual classifiers. The SVM sequence had slightly better accuracy than the SVM committee but also had bigger
We evaluate and improve an existing curvature-based region growing algorithm for colonic polyp detection for our CT colonography (CTC) computer-aided detection (CAD) system by using Pareto fronts. The performance of a polyp detection algorithm involves two conflicting objectives, minimizing both false negative (FN) and false positive (FP) detection rates. This problem does not produce a single optimal solution but a set of solutions known as a Pareto front. Any solution in a Pareto front can only outperform other solutions in one of the two competing objectives. Using evolutionary algorithms to find the Pareto fronts for multi-objective optimization problems has been common practice for years. However, they are rarely investigated in any CTC CAD system because the computation cost is inherently expensive. To circumvent this problem, we have developed a parallel program implemented on a Linux cluster environment. A data set of 56 CTC colon surfaces with 87 proven positive detections of polyps sized 4 to 60 mm is used to evaluate an existing one-step, and derive a new two-step region growing algorithm. We use a popular algorithm, the Strength Pareto Evolutionary Algorithm (SPEA2), to find the Pareto fronts. The performance differences are evaluated using a statistical approach. The new algorithm outperforms the old one in 81.6% of the sampled Pareto fronts from 20 simulations. When operated at a suitable sensitivity level such as 90.8% (79/87) or 88.5% (77/87), the FP rate is decreased by 24.4% or 45.8% respectively.
We present a hybrid committee classifier for computer-aided detection (CAD) of colonic polyps in CT colonography (CTC). The classifier involved an ensemble of support vector machines (SVM) and neural networks (NN) for classification, a progressive search algorithm for selecting a set of features used by the SVMs and a floating search algorithm for selecting features used by the NNs. A total of 102 quantitative features were calculated for each polyp candidate found by a prototype CAD system. 3 features were selected for each of 7 SVM classifiers which were then combined to form a committee of SVMs classifier. Similarly, features (numbers varied from 10-20) were selected for 11 NN classifiers which were again combined to form a NN committee classifier. Finally, a hybrid committee classifier was defined by combining the outputs of both the SVM and NN committees. The method was tested on CTC scans (supine and prone views) of 29 patients, in terms of the partial area under a free response receiving operation characteristic (FROC) curve (AUC). Our results showed that the hybrid committee classifier performed the best for the prone scans and was comparable to other classifiers for the supine scans.
Colonic polyps appear like elliptical protrusions on the inner wall of the colon. Curvature based features for colonic polyp detection have proved to be successful in several computer-aided diagnostic CT colonography (CTC) systems. Some simple thresholds are set for those features for creating initial polyp candidates, sophisticated classification scheme are then applied on these polyp candidates to reduce false positives. There are two objective functions, the number of missed polyps and false positive rate, that need to be minimized when setting those thresholds. These two objectives conflict and it is usually difficult to optimize them both by a gradient search. In this paper, we utilized a multiobjective evolutionary method, the Strength Pareto Evolutionary Algorithm (SPEA2), to optimize those thresholds. SPEA2 incorporates the concept of Pareto dominance and applies genetic techniques to evolve individual solutions to the Pareto front. The SPEA2 algorithm was applied to colon CT images from 27 patients each having a prone and a supine scan. There are 40 colonoscopically confirmed polyps resulting in 72 positive detections in CTC reading. The results obtained by SPEA2 were compared with those obtained by our old system, where an appropriate value was set for each of those thresholds by a histogram examination method. If we keep the sensitivity the same as that of our old system, the SPEA2 algorithm reduced false positive rate by 76.4% from average false positive 55.6 to 13.3 per data set. If the false positive rate is kept the same for both systems, SPEA2 increased the sensitivity by 13.1% from 53 to 61 among 72 ground truth detections.
The computed tomographic colonography (CTC) computer aided detection (CAD) program is a new method in development to detect colon polyps in virtual colonoscopy. While high sensitivity is consistently achieved, additional features are desired to increase specificity. In this paper, a wavelet analysis was applied to CTCCAD outputs in an attempt to filter out false positive detections.
52 CTCCAD detection images were obtained using a screen capture application. 26 of these images were real polyps, confirmed by optical colonoscopy and 26 were false positive detections. A discrete wavelet transform of each image was computed with the MATLAB wavelet toolbox using the Haar wavelet at levels 1-5 in the horizontal, vertical and diagonal directions. From the resulting wavelet coefficients at levels 1-3 for all directions, a 72 feature vector was obtained for each image, consisting of descriptive statistics such as mean, variance, skew, and kurtosis at each level and orientation, as well as error statistics based on a linear predictor of neighboring wavelet coefficients. The vectors for each of the 52 images were then run through a support vector machine (SVM) classifier using ten-fold cross-validation training to determine its efficiency in distinguishing polyps from false positives.
The SVM results showed 100% sensitivity and 51% specificity in correctly identifying the status of detections. If this technique were added to the filtering process of the CTCCAD polyp detection scheme, the number of false positive results could be reduced significantly.