Research On Medical Image Classification Algorithm Based On Self-Supervised Learning

With the acceleration of the construction of smart medical informatization,more and more attention has been paid to computer-aided diagnosis technology.As an important method of computer-aided diagnosis technology,deep learning has great development potential in the field of medical image classification.However,the training of deep learning models re-quires a large amount of data to support,and complete labeled datasets in the field of medical image classification are very scarce.The transfer learning method is used to deal with the problem of data shortage and achieve good results,but the difference between natural images and medical images cannot be ignored,and this difference will also affect the classification accuracy of the model.Self-supervised learning avoids this problem very well.Only a small amount of labeled data and a large amount of unlabeled data are required to allow the model to obtain a good classification accuracy.Self-supervised learning uses a large amount of unlabeled data for model pre-training,and then fine-tunes it on a small amount of labeled data,which greatly reduces the dependence of deep learning methods on labeled data.How-ever,the current research and development of self-supervised learning methods in the field of medical image classification is not yet mature,and there is still room for improvement.Therefore,aiming at the scarcity of labeled medical image classification datasets,this paper conducts in-depth research on self-supervised learning methods.The main research contents are as follows:1.This paper proposes a novel coronavirus pneumonia detection algorithm based on hard negative sample contrastive learning(Fast Mixing Contrastive Learning,FMix CL).In clin-ical acquisition of medical images,the imaging results are greatly unstable due to lighting environment,equipment parameters,patient factors,etc.,and the model can learn transfor-mation invariance through multiple image augmentation methods.Hard negative samples in contrastive learning are the key factors for the model to learn more robust visual features.This paper introduces a module for fastly mixing hard negative samples.The query feature samples and the existing hard negative samples are linearly convexly combined to synthe-size more hard negative samples into the model training process.By increasing the difficulty of the self-supervised pretext task,the proposed method promotes the model to learn more differentiated features,and enables the model to have stronger discrimination ability for med-ical images containing different semantic information,thus achieving higher classification accuracy in COVID-19 classification tasks.2.This paper proposes a generative self-supervised learning method based on Transformer(Masked Auto Encoder-Conv-Transformer,MAE-Cv T)and applies it to the task of skin can-cer classification.Medical images all come from the same part or tissue,and the images of different patients have high similarity,while the difference is often reflected in the local lesion site at the pixel level.MAE-Cv T blocks part of the image pixels and then restores them.Such pretext task setting mode makes the model pay more attention to the pixel-level information of medical images,so as to extract more different image features from similar medical images.The MAE-Cv T method evenly divides the image into multiple image blocks and randomly masks the image blocks at a ratio of 81.5%.Only the unmasked image blocks are sent to the Encoder to extract the image code,and then the Decoder is used to decode the image code to restore the image.This pretext task allows the Encoder to acquire the ability to extract excellent image features.In this paper,an improved Con V-Transformer model is proposed,in which the convolution module is added to improve the local modeling ability,and the Embedding length is reduced and the number of model parameters is reduced greatly.Conv-transformer,as the Encoder of MAe-CVT,enhances its ability of modeling local in-formation of image,and promotes it to extract image features faster and better.The Encoder trained by MAE-CVT was extracted for fine-tuning,and it had good classification accuracy in skin cancer classification task.