Xiangyu Si's research while affiliated with Xi'an Jiaotong University and other places

Image segmentation for human organs is an important task for diagnosis and treatment of diseases. Current deep learning-based methods are fully supervised that need pixel-level labels. Since the medical images are highly specialized and complex, the work of delineating pixel-level segmentation masks is very time-consuming. Weakly supervised methods are then chosen to lighten the workload, which only needs physicians to determine whether an image contains the organ regions of interest. While these weakly supervised methods have a common drawback. They do not incorporate prior knowledge that alleviates the lack of pixel-level information for segmentation. In this work, we propose a weakly supervised method based on prior knowledge for the segmentation of human organs. The proposed method was validated on three data sets of human organ segmentation. Experimental results show that the proposed image-level supervised segmentation method outperforms several state-of-the-art methods.

Level set method is widely used for image segmentation. Recent work combined traditional level set method with deep learning architecture for image segmentation. However, it is limited when dealing with medical images because of the blurred edges and complex intensity distribution, which leads to the loss of spatial details. To address this problem, we propose a deep level set method to refine object boundary details and improve the segmentation accuracy. We integrate augmented prior knowledge into inputs of CNN, which can make the level set evolution result has more accurate shape. In addition, to consider the spatial correlation of the object, we combine a boundary loss with deep level set model for preventing the reduction of details. We evaluate the proposed method on two medical image data sets, which are prostate magnetic resonance images and retinal fundus images. The experimental results show that the proposed method achieves state-of-the-art performance.

Image segmentation technology has made a remarkable effect in medical image analysis and processing, which is used to help physicians get a more accurate diagnosis. Manual segmentation of the medical image requires a lot of effort by professionals, which is also a subjective task. Therefore, developing an advanced segmentation method is an essential demand. We propose an end-to-end segmentation method for medical images, which mimics physicians delineating a region of interest (ROI) on the medical image in a multi-step manner. This multi-step operation improves the performance from a coarse result to a fine result progressively. In this paper, the segmentation process is formulated as a Markov decision process and solved by a deep reinforcement learning (DRL) algorithm, which trains an agent for segmenting ROI in images. The agent performs a serial action to delineate the ROI. We define the action as a set of continuous parameters. Then, we adopted a DRL algorithm called deep deterministic policy gradient to learn the segmentation model in continuous action space. The experimental result shows that the proposed method has 7.24% improved to the state-of-the-art method on three prostate MR data sets and has 3.52% improved on one retinal fundus image data set.

Style transfer is an increasingly popular field that can capture the styles of a particular artwork and use them to synthesize a new image with specific content. Previous NST algorithms have the limitation to transfer styles to correct regions in the output image. Therefore, some regions in the output image have deformed structures of the source image. In this paper, we propose a color preprocessing-based neural style transfer method to overcome the limitation. To reduce impacts caused by color differences between source image and style, we propose three models based on a color iterative distribution transform algorithm (IDT). The first one is named original color-preprocessed (OCp) model, which uses IDT to transform the color probability density function (PDF) of source image into that of style image. The second one is named exposure-corrected original color-preprocessed (EC-OCp) model, which adds an automatic detail-enhanced exposure correction module before OCp model. When source image is underexposed, EC-OCp model can achieve better results than OCp model. The third one is style color-preprocessed (SCp) model. It uses IDT to transform the color PDF of style image into that of source image. The original structures are well protected in the output image. According to experiments, the proposed models are robust to the source images with more conditions. Therefore, they have more usage values than the original method.