Research On Key Technologies Of Medical Image Recognition Based On Context And Multi-task Learning

In the development of modern medicine,medical images including histopathology,MRI and computer tomography have become one of the important diagnostic evidences that are indispensable in clinical medical data.However,in the face of increasing medical demand and the accompanying large amount of medical imaging data,doctors need to diagnose and screen thousands of films every day.This too mechanical and repetitive process will consume a lot of doctors’ time and energy.And this kind of long-time and high-intensity work will inevitably cause some diagnostic errors.Therefore,the introduction of medical image recognition technology with the help of artificial intelligence and deep learning technology is particularly important,and whether it can help doctors to give accurate diagnosis results in a relatively short period of time has become the one of the test standards for whether the technology can be truly implemented.This thesis focuses on trying to solve two common problems for medical image recognition algorithm based on deep learning.One is that most medical image data is difficult to label,so that there is often a lack of data,which affects the generalization ability of the model;the other is that the original medical image size is so large that it is difficult to directly input the model,so the image needs to be sliced before entering the model.This causes the problem of spatial information loss,and affecting the accuracy of the model’s estimation.In order to solve the above problems.first of all,to address the problem of airspace information loss,we proposed a general LSTM-based basic airspace relationship model,and an acceleration method based on feature sharing is designed to optimize the model training and the efficiency problems.Secondly,to solve the problem of lack of medical image data,a semi-supervised multi-task learning method is proposed.Through the design of auxiliary tasks,unlabeled samples can be effectively used for training.Finally,this paper combines the two ideas of multi-task learning and proposes a general algorithm for the recognition of medical image lesions to maximize the model’s learning ability and prediction accuracy and generalization ability.