Recent progress in deep learning, especially deep convolutional neural networks (DCNNs), has led to significant improvement in natural image classification. However, research is still ongoing in the domain of medical image analysis in part due to the shortage of annotated data sets for training DCNNs, the imbalanced number of positive and negative samples, and the difference between medical images and natural images. In this paper, two strategies are proposed to train a DCNN for pathological lymph node image classification. Firstly, the transfer learning strategy is used to deal with the shortage of training samples. Second, a novel exponential loss function is presented for the imbalance in training samples. Four state-of-the-art DCNNs (GoogleNet, ResNet101, Xception, and MobileNetv2) are tested. The experiments demonstrate that the two strategies are effective to improve the performance of pathological lymph node image classification in terms of accuracy and sensitivity with a mean of 0.13% and 1.50%, respectively, for the four DCNNs. In particular, the proposed exponential loss function improved the sensitivity by 3.9% and 4.0% for Xception and ResNet101, respectively.
As the digital finite impulse response (FIR) filter can be transformed into the shift-add form of multiple small-sized firstorder moments, based on the existing fast first-order moment algorithm, this paper presents a novel multiplier-less structure to calculate any number of sequential filtering results in parallel. The theoretical analysis on its hardware and time-complexities reveals that by appropriately setting the degree of parallelism and the decomposition factor of a fixed word width, the proposed structure may achieve better area-time efficiency than the existing two-dimensional (2-D) memoryless-based filter. To evaluate the performance concretely, the proposed designs for different taps along with the existing 2-D memoryless-based filters, are synthesized by Synopsys Design Compiler with 0.18-μm SMIC library. The comparisons show that the proposed design has less area-time complexity and power consumption when the number of filter taps is larger than 48.
Lesion segmentation has remained a challenge in different body regions. Generalizability is lacking in published methods as variability in results is common, even for a given organ and modality, such that it becomes difficult to establish standardized methods of disease quantification and reporting. This paper makes an attempt at a generalizable method based on classifying lesions along with their background into groups using clinically used visual attributes. Using an Iterative Relative Fuzzy Connectedness (IRFC) delineation engine, the ideas are implemented for the task of liver lesion segmentation in computed tomography (CT) images. For lesion groups with the same background properties, a few subjects are chosen as the training set to obtain the optimal IRFC parameters for the background tissue components. For lesion groups with similar foreground properties, optimal foreground parameters for IRFC are set as the median intensity value of the training lesion subset. To segment liver lesions belonging to a certain group, the devised method requires manual loading of the corresponding parameters, and correct setting of the foreground and background seeds. The segmentation is then completed in seconds. Segmentation accuracy and repeatability with respect to seed specification are evaluated. Accuracy is assessed by the assignment of a delineation quality score (DQS) to each case. Inter-operator repeatability is assessed by the difference between segmentations carried out independently by two operators. Experiments on 80 liver lesion cases show that the proposed method achieves a mean DQS score of 4.03 and inter-operator repeatability of 92.3%.
This paper presents hardware efficient designs for implementing the one-dimensional (1D) discrete Fourier transform (DFT). Once DFT is formulated as the cyclic convolution form, the improved first-order moments-based cyclic convolution structure can be used as the basic computing unit for the DFT computation, which only contains a control module, a barrel shifter and (N-1)/2 accumulation units. After decomposing and reordering the twiddle factors, all that remains to do is shifting the input data sequence and accumulating them under the control of the statistical results on the twiddle factors. The whole calculation process only contains shift operations and additions with no need for multipliers and large memory. Compared with the previous first-order moments-based structure for DFT, the proposed designs have the advantages of less hardware consumption, lower power consumption and the flexibility to achieve better performance in certain cases. A series of experiments have proven the high performance of the proposed designs in terms of the area time product and power consumption. Similar efficient designs can be obtained for other computations, such as DCT/IDCT, DST/IDST, digital filter and correlation by transforming them into the forms of the first-order moments based cyclic convolution.
In this paper, we proposed a novel method to extract shape feature based on dual-tree complex wavelet. First, with the two level dual-tree complex wavelet transformations, we can get two low frequency components of the first level, which are used as wavelet moment invariants formed from approximation coefficients. Then, we calculate means and variance for each of the six detailed components in the second level since it contains different directions information of the shape. Using the Principal Component Analysis (PCA), twenty features can be reduced to five maximum useful features which contribute to shape matching.
Proc. SPIE. 8920, MIPPR 2013: Parallel Processing of Images and Optimization and Medical Imaging Processing
KEYWORDS: Statistical analysis, Clocks, Digital filtering, Fourier transforms, Data processing, Very large scale integration, Information technology, Convolution, Algorithm development, Computer architecture
This paper presents a hardware-efficient design for the one-dimensional (1-D) discrete Fourier transform (DFT). Once
the 1-D DFT is formulated as the cyclic convolution form, the first-order moments-based structure can be used as the
basic computing unit for the DFT computation, which only contains a control module, a statistical module and an
accumulation module. The whole calculation process only contains shift operations and additions, with no need for
multipliers and large memory. Compared with the traditional DA-based structure for DFT, the proposed design has better
performance in terms of the area-throughput ratio and the power consumption, especially when the length of DFT is
slightly longer. Similar efficient designs can be obtained for other computations, such as the DCT/IDCT, DST/IDST,
digital filter and correlation, by transforming them into the forms of the first-order moments-based cyclic convolution.