Image Segmentation Algorithm For Weakly Annotated Gastric Cancer Slices Based On Deep Learning

Cancer is an incurable disease for human beings,among which gastric cancer with its high incidence and high mortality rate ranks second in China.Since the middle and late 20th century,researchers began to study the relevant techniques of medical image segmentation to alleviate the heavy burden of pathologists.However,these algorithms are usually less robust,complex and cannot provide much semantic information.Meanwhile,the high cost of manual labeling also limits the application of supervised learning in the field of medical image segmentation.Therefore,the application of semi-supervised deep learning in the diagnosis of gastric cancer can not only save the lives of countless patients,but also has great significance in relieving medical resources,saving costs and conflicts between doctors and patients.Based on the analysis of the research and application of deep learning in medical image recognition at home and abroad,this paper makes a detailed study on the recognition of cancer pathology section.By combining traditional image segmentation based on patches with fully convolutional neural network,this paper explores image semantic segmentation based on weakly labeled data sets through semi-supervised learning.In order to further enhance the generalization ability and performance of the model,an overlapped region forecast algorithm is proposed in this paper.To further improve the capability of the model,the paper uses a fully connected conditional random field as post-processing at the end of the network.By calculating the relationship between the probability map of the output and the features of the color texture between pixels in the original image,similar pixels are assigned the same label.This paper also proposes a learning mechanism of reiterative learning,which can improve the performance and precision of the model by predicting and processing the training set through the existing model.In addition,in view of the shortcomings of current open source data labeling software such as poor user experience and tedious labeling process,we developed a data labeling software in order to achieve high efficiency and shortcut data labeling.By training convolutional neural networks based on patches,preprocessing with digital image technics,post-processing with full connected conditional random field,doing overlapped region forecast and other algorithms,this paper promotes the effect with the learning mechanism of reiterative learning.Our model achieved a mean Intersection over Union coefficient(IOU)of 88.31,an average accuracy of 91.09%without any subsequent manual labeling,and won the 2017 China Big Data Artificial Intelligence Innovation and Entrepreneurship Competition.Through this paper,it is proved that semi-supervised learning on weakly labeled images can achieve near supervised learning effect.This study greatly reduces the dependence of the model on the gold standard dataset and saves the cost,and is of great significance to the development of semantic segmentation of medical imagesThe primary contributions of this paper are as follows:1.An advisable training framework,reiterative learning,is presented to train our network on a partially annotated biomedical dataset and achieves state-of-the-art performance without pre-training or further manual annotation.2.We combine the patch-based approach with an FCN-based model and present a novel method to solve gastric tumor segmentation problems.3.An overlapped region forecast algorithm is proposed to merge the predictions and then promote the performance of the final results.4.To the best of our knowledge,we are the first group to use deep learning to solve pathological image segmentation problems for gastric cancer.