A Study Of Medical Image Classification Algorithms Based On Self-supervised Representation Learning

In recent years,deep learning-based models have demonstrated excellent algorithmic performance in many fields,including vision,text and audio.However,designing machine learning algorithms with insufficient data is still a major challenge in computer vision industrial.To solve this problem,the research topic of few-shot learning had been proposed and gradually becoming a research hotspot in the field of machine learning.The difficulty of collecting medical images,which is often due to several reasons such as the data object is a sudden disease or infrequent special symptoms,and data labels need to be diagnosed by medical experts.Therefore,the practical classification algorithms for insufficient medical data was took as a research purpose for this thesis and few-shot learning and meta-learning of complex objectives was took as main research content.This thesis first defines the problem of machine learning with insufficient data,and explains the fundamental reason why deep model have difficult to popularize in this case.By analyzing the design shortcomings of current mainstream methods,this thesis proposes that the poor classification results of mainstream methods on natural images are caused by the insufficient feature description ability of the model.This thesis design a self-supervised representation learning algorithm on medical data and enhance the fewshot model with that aforementioned representation encoding to achieve higher performance in medical image classification.The details are as follows.(1)This thesis analyzes the contrast predicting coding technique and the generalized visual representation coding algorithm Sim CLR,investigate contrast learning methods and pretext tasks of both.By improving self-supervised training process,a multi-step contrast prediction model and global constrastive visual representation algorithm are proposed.Applying the popular semi-supervised verification protocol,this thesis constructs a verification experiment of the above representation coding modules.On medical image data,our proposal representation coding methods brings more than 10% classification accuracy improvement for the downstream supervised training with Res Net,which demonstrates the usability of the representation coding module this thesis proposed.(2)The downstream few-shot learning task is reinforced by the form of representational encoding,this thesis discusses the representational few-shot learning algorithm and the meta-learning algorithm respectively.After migration of the representation encoder and model customization,a few-shot learning algorithm based on multi-step contrast prediction enhancement and a meta-learning algorithm based on global constrastive enhancement are proposed.The proposal algorithms have achieve more than 15% classification accuracy improvement compared to the native algorithms on the dermoscopic image dataset ISIC2018,and OCTMNIST,and Tissue MNIST,from the medical databases Med MNIST which are difficult to classify.In the model convergence analysis,our proposal algorithms are able to converge faster and exhibit a more stable training trajectory compared to the native few-shot learning algorithm.This implies that the representation coding model can clustering the input into a shallow learning effective feature space,which enables the downstream few-shot learning algorithm to achieve better classification performance with the reinforcement of the representation coding.