Deep Learning And Polar Representation Based Medical Image Semantic Segmentation

Modern clinical diagnosis highly depends on medical imaging techniques.Through the medical images,doctors can confirm the diseased areas of the human internal organs or tissues,and then put forward the diagnosis and treatment correspondingly.However,since the anatomical structures of the organs and tissues are complex,even an experienced doctor costs a certain amount of time and energy to interpret medical images.Since the medical image segmentation technique can locate the diseased areas quickly and accurately,and then help improve the diagnosis efficiency,it has been widely concerned.With the great success of the convolutional neural networks achieved in machine vision,the accuracy and robustness of the medical image segmentation technique have been improved greatly.While there remain a lot of challenges.Medical images usually contain lots of noise and artifacts,and the contours of the targets are fuzzy,making the accurate segmentation extremely hard.Furthermore,the sources of medial images are limited,and the high-quality segmentation labels depend on the manual annotations of professional doctors which are costly.The network is hard to be trained with limited data.For the above challenges,this paper takes the common circular and oval-shaped targets as the main research objects.A polar regression based fully-supervised segmentation network and a manifold based semi-supervised segmentation network are proposed then.By making full use of the anatomical information and relieving the dependency of labeled data,both networks achieve good segmentation results with limited training data.The main work is as follows:Polar regression based fully-supervised segmentation network: For the complex representation of the target’s shape in traditional segmentation networks,this paper proposes a fully-supervised network based on the polar representation.The network represents the target as a center point and a set of polar rays,and realizes segmentation with the regression of center map and polar rays to be more shape aware.To generate the ground truth of the center map,we propose a contour-based algorithm.In the polar ray length regression,the center map is applied as a mask of the loss,which emphasizes the importance of pixels with high center probability to polar ray regression.Besides,we construct the length constraint of polar ray pairs and puts forward a projection distance loss accordingly.It helps make full use of the shape and size information of the target in polar maps,and further advances the convergence of the network.Several experiments are carried out on 3 public medical image datasets,and the effectiveness of the network is proved by comparing to other similar networks.Manifold based semi-supervised segmentation network: Aiming at the problem of sparsely labeled data,this paper further proposes a manifold based semi-supervised network referring to the idea of manifold embedding.We design a low dimensional manifold label to represent the shape and size characteristics of the target,and appends two manifold branches to the network accordingly.These two branches take the original image and the predicted polar ray length map as inputs and predict the manifold maps.So they can construct the semi-supervised signal with the consistency requirement in the manifold space.To make proper use of the unlabeled data,the network applies a two-stage training strategy.In detail,the network parameters are initialized by fully-supervised training with the labeled data first,and then optimized by semi-supervised training with both labeled and unlabeled data in a certain proportion.To prove the effectiveness,we conduct a series of ablation experiments and comparative experiments on 2 public medical image segmentation datasets.The results show that with 20% training data labeled,our method is superior to other semi-supervised segmentation methods on Dice and HD（Hausdorff Distance）score.