Electrical Engineering And Systems Science

Intervertebral discs (IVDs), as small joints lying between adjacent vertebrae, have played an important role in pressure buffering and tissue protection. The fully-automatic localization and segmentation of IVDs have been discussed in the literature for many years since they are crucial to spine disease diagnosis and provide quantitative parameters in the treatment. Traditionally hand-crafted features are derived based on image intensities and shape priors to localize and segment IVDs. With the advance of deep learning, various neural network models have gained great success in image analysis including the recognition of intervertebral discs. Particularly, U-Net stands out among other approaches due to its outstanding performance on biomedical images with a relatively small set of training data. This paper proposes a novel convolutional framework based on 3D U-Net to segment IVDs from multi-modality MRI images. We first localize the centers of intervertebral discs in each spine sample and then train the network based on the cropped small volumes centered at the localized intervertebral discs. A detailed comprehensive analysis of the results using various combinations of multi-modalities is presented. Furthermore, experiments conducted on 2D and 3D U-Nets with augmented and non-augmented datasets are demonstrated and compared in terms of Dice coefficient and Hausdorff distance. Our method has proved to be effective with a mean segmentation Dice coefficient of 89.0% and a standard deviation of 1.4%.

Age related macular degeneration (AMD) is one of the major causes for blindness in the elderly population. In this report, we propose deep learning based methods for retinal analysis using color fundus images for computer aided diagnosis of AMD. We leverage the recent state of the art deep networks for building a single fundus image based AMD classification pipeline. We also propose methods for the other directly relevant and auxiliary tasks such as lesions detection and segmentation, fovea detection and optic disc segmentation. We propose the use of generative adversarial networks (GANs) for the tasks of segmentation and detection. We also propose a novel method of fovea detection using GANs.

The coronavirus (COVID-19) is currently the most common contagious disease which is prevalent all over the world. The main challenge of this disease is the primary diagnosis to prevent secondary infections and its spread from one person to another. Therefore, it is essential to use an automatic diagnosis system along with clinical procedures for the rapid diagnosis of COVID-19 to prevent its spread. Artificial intelligence techniques using computed tomography (CT) images of the lungs and chest radiography have the potential to obtain high diagnostic performance for Covid-19 diagnosis. In this study, a fusion of convolutional neural network (CNN), support vector machine (SVM), and Sobel filter is proposed to detect COVID-19 using X-ray images. A new X-ray image dataset was collected and subjected to high pass filter using a Sobel filter to obtain the edges of the images. Then these images are fed to CNN deep learning model followed by SVM classifier with ten-fold cross validation strategy. This method is designed so that it can learn with not many data. Our results show that the proposed CNN-SVM with Sobel filtering (CNN-SVM+Sobel) achieved the highest classification accuracy of 99.02% in accurate detection of COVID-19. It showed that using Sobel filter can improve the performance of CNN. Unlike most of the other researches, this method does not use a pre-trained network. We have also validated our developed model using six public databases and obtained the highest performance. Hence, our developed model is ready for clinical application

k-space undersampling is a standard technique to accelerate MR image acquisitions. Reconstruction techniques including GeneRalized Autocalibrating Partial Parallel Acquisition(GRAPPA) and its variants are utilized extensively in clinical and research settings. A reconstruction model combining GRAPPA with a conditional generative adversarial network (GAN) was developed and tested on multi-coil human brain images from the fastMRI dataset. For various acceleration rates, GAN and GRAPPA reconstructions were compared in terms of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). For an acceleration rate of R=4, PSNR improved from 33.88 using regularized GRAPPA to 37.65 using GAN. GAN consistently outperformed GRAPPA for various acceleration rates.

Catchy but rigorous deep learning architectures were tailored for image super-resolution (SR), however, these fail to generalize to non-Euclidean data such as brain connectomes. Specifically, building generative models for super-resolving a low-resolution (LR) brain connectome at a higher resolution (HR) (i.e., adding new graph nodes/edges) remains unexplored although this would circumvent the need for costly data collection and manual labelling of anatomical brain regions (i.e. parcellation). To fill this gap, we introduce GSR-Net (Graph Super-Resolution Network), the first super-resolution framework operating on graph-structured data that generates high-resolution brain graphs from low-resolution graphs. First, we adopt a U-Net like architecture based on graph convolution, pooling and unpooling operations specific to non-Euclidean data. However, unlike conventional U-Nets where graph nodes represent samples and node features are mapped to a low-dimensional space (encoding and decoding node attributes or sample features), our GSR-Net operates directly on a single connectome: a fully connected graph where conventionally, a node denotes a brain region, nodes have no features, and edge weights denote brain connectivity strength between two regions of interest (ROIs). In the absence of original node features, we initially assign identity feature vectors to each brain ROI (node) and then leverage the learned local receptive fields to learn node feature representations. Second, inspired by spectral theory, we break the symmetry of the U-Net architecture by topping it up with a graph super-resolution (GSR) layer and two graph convolutional network layers to predict a HR graph while preserving the characteristics of the LR input. Our proposed GSR-Net framework outperformed its variants for predicting high-resolution brain functional connectomes from low-resolution connectomes.

Deep segmentation models that generalize to images with unknown appearance are important for real-world medical image analysis. Retraining models leads to high latency and complex pipelines, which are impractical in clinical settings. The situation becomes more severe for ultrasound image analysis because of their large appearance shifts. In this paper, we propose a novel method for robust segmentation under unknown appearance shifts. Our contribution is three-fold. First, we advance a one-stage plug-and-play solution by embedding hierarchical style transfer units into a segmentation architecture. Our solution can remove appearance shifts and perform segmentation simultaneously. Second, we adopt Dynamic Instance Normalization to conduct precise and dynamic style transfer in a learnable manner, rather than previously fixed style normalization. Third, our solution is fast and lightweight for routine clinical adoption. Given 400*400 image input, our solution only needs an additional 0.2ms and 1.92M FLOPs to handle appearance shifts compared to the baseline pipeline. Extensive experiments are conducted on a large dataset from three vendors demonstrate our proposed method enhances the robustness of deep segmentation models.

The shortage of annotated medical images is one of the biggest challenges in the field of medical image computing. Without a sufficient number of training samples, deep learning based models are very likely to suffer from over-fitting problem. The common solution is image manipulation such as image rotation, cropping, or resizing. Those methods can help relieve the over-fitting problem as more training samples are introduced. However, they do not really introduce new images with additional information and may lead to data leakage as the test set may contain similar samples which appear in the training set. To address this challenge, we propose to generate diverse images with generative adversarial network. In this paper, we develop a novel generative method named generative adversarial U-Net , which utilizes both generative adversarial network and U-Net. Different from existing approaches, our newly designed model is domain-free and generalizable to various medical images. Extensive experiments are conducted over eight diverse datasets including computed tomography (CT) scan, pathology, X-ray, etc. The visualization and quantitative results demonstrate the efficacy and good generalization of the proposed method on generating a wide array of high-quality medical images.

Research in vertebral bone micro-structure generally requires costly procedures to obtain physical scans of real bone with a specific pathology under study, since no methods are available yet to generate realistic bone structures in-silico. Here we propose to apply recent advances in generative adversarial networks (GANs) to develop such a method. We adapted style-transfer techniques, which have been largely used in other contexts, in order to transfer style between image pairs while preserving its informational content. In a first step, we trained a volumetric generative model in a progressive manner using a Wasserstein objective and gradient penalty (PWGAN-GP) to create patches of realistic bone structure in-silico. The training set contained 7660 purely spongeous bone samples from twelve human vertebrae (T12 or L1) with isotropic resolution of 164um and scanned with a high resolution peripheral quantitative CT (Scanco XCT). After training, we generated new samples with tailored micro-structure properties by optimizing a vector z in the learned latent space. To solve this optimization problem, we formulated a differentiable goal function that leads to valid samples while compromising the appearance (content) with target 3D properties (style). Properties of the learned latent space effectively matched the data distribution. Furthermore, we were able to simulate the resulting bone structure after deterioration or treatment effects of osteoporosis therapies based only on expected changes of micro-structural parameters. Our method allows to generate a virtually infinite number of patches of realistic bone micro-structure, and thereby likely serves for the development of bone-biomarkers and to simulate bone therapies in advance.

We introduce a novel unsupervised deep generative manifold model for the recovery of multi-slice free-breathing and ungated cardiac MRI from highly undersampled measurements. The proposed scheme represents the multi-slice volume at each time point as the output of a deep convolutional neural network (CNN) generator, which is driven by latent vectors that capture the cardiac and respiratory phase at the specific time point. The main difference between the proposed method and the traditional CNN approaches is that the proposed scheme learns the network parameters from only the highly undersampled data rather than the extensive fully-sampled training data. We also learn the latent codes from the undersampled data using the stochastic gradient descent. Regularizations on the network and the latent codes are introduced to encourage the learning of smooth image manifold and the latent codes for each slice have the same distribution. The main benefits of the proposed scheme are (a) the ability to align multi-slice data and capitalize on the redundancy between the slices; (b) the ability to estimate the gating information directly from the k-t space data; and (c) the unsupervised learning strategy that eliminates the need for extensive training data.

Modern single image super-resolution (SISR) system based on convolutional neural networks (CNNs) achieves fancy performance while requires huge computational costs. The problem on feature redundancy is well studied in visual recognition task, but rarely discussed in SISR. Based on the observation that many features in SISR models are also similar to each other, we propose to use shift operation to generate the redundant features (i.e., Ghost features). Compared with depth-wise convolution which is not friendly to GPUs or NPUs, shift operation can bring practical inference acceleration for CNNs on common hardware. We analyze the benefits of shift operation for SISR and make the shift orientation learnable based on Gumbel-Softmax trick. For a given pre-trained model, we first cluster all filters in each convolutional layer to identify the intrinsic ones for generating intrinsic features. Ghost features will be derived by moving these intrinsic features along a specific orientation. The complete output features are constructed by concatenating the intrinsic and ghost features together. Extensive experiments on several benchmark models and datasets demonstrate that both the non-compact and lightweight SISR models embedded in our proposed module can achieve comparable performance to that of their baselines with large reduction of parameters, FLOPs and GPU latency. For instance, we reduce the parameters by 47%, FLOPs by 46% and GPU latency by 41% of EDSR x2 network without significant performance degradation.