Medical Image Recognition Via Deep Learning

The development of deep learning technology brings opportunities and challenges to medical image recognition.At present,deep learning technology has been applied to the diagnosis of lung CT,diabetic retinopathy image,etc.,but for the second highest incidence of gastric cancer,the current research is still in a blank stage.In this paper,We delve into the relevant methods of deep learning in the field of image recognition,and designed a high performance deep learning algorithm which can be used for the classification and segmentation of Gastric cancer pathological sections by combining the characteristics of pathological section data of gastric cancer.The model trained in this paper can judging whether the pathological section of gastric cancer is ill and then segment the fine contour of the cancerous region in the pathological section of gastric cancer with a very high accuracy rate,where the clasification score is 96.67%and the segmentation score reached 86.87%.The research work in this paper is helpful to improve the efficiency and accuracy of gastric cancer detection,playing the goal of assisting doctors in diagnosis and treatment,and has far-reaching significance for saving lives and exploring the mysteries of life and health.This paper mainly has the following innovations and contribution:(1)In this paper,a deep convolution network structure based on multi-task learning and multi-feature fusion is proposed.On the one hand,the multi-task learning method combines classification and segmentation tasks,and can simultaneously judge the Negative and Positive of gastric Cancer pathological sections and segment the contour of cancerous region.On the other hand,the multi-feature fusion method can effectively improve the performance of the model by combining the information of different receptive field and different scales.(2)In this paper,we present a series of methods to improve the performance of a single model.First of all,we proposed two important improvements:mask learning and traversal search for segmentation threshold.Secondly,according to the character-istics of the dataset and task of the pathological section of gastric cancer,a very effective data enhancement method,such as elastic deformation,is proposed,and the output prediction map is treated by morphology.Finally,we analyzed the ensemble methods in-depth and proposed a series of multi-model fusion methods to further improve the performance of the network.The algorithm in this paper won the fifth place in the final of BOT gastric cancer pathological section recognition AI challenge.(3)In this paper,we introduce some details encountered in the course of the experiment,including hyper-parameter setting,algorithm tuning strategy and so on.More importantly,a lot of comparative and validation experiments are shown in this paper,which fully verified the effectiveness of each algorithm module and provide an effective reference for other researchers.