Analysis Of Different Deep Learning Models On Medical Images

Image classification is a core task in the computer vision field,from the traditional machine learning to the current deep learning,the mainstream deep learning models are mainly studied on natural images,however,medical images are very different from natural images in terms of imaging principle,data volume,data quality,feature expression,etc.If deep learning models can be effectively used for medical image classification,learning and recognizing complex features in medical images,it can help doctors improve the accuracy and timeliness of diagnosis,accelerate the diagnosis and treatment of diseases,reduce medical costs thus promoting the further development of intelligent medical care.The purpose of thesis is to investigate the application of different deep learning models on medical images,and to construct a novel interpretable Multi-Layer Perceptron(MLP)network.The main research components are as follows:(1)For four widely used different deep learning models,this thesis migrates their processing of natural images to the medical image classification task and compare their performance.The similarities and differences between natural images and medical images in terms of feature information and data quality are analyzed,and the feasibility of different deep learning model architectures on medical image classification tasks is experimentally explored.And based on the experimental results of different deep learning model architectures on natural image and medical image datasets,our thesis further investigates the improvement directions of combining the existing deep learning models with medical image classification tasks.(2)To address the problem that medical image classification focuses more on finding key lesion regions and the poor interpretability of traditional deep neural networks,this thesis designs a new Sparse-attention based Residual Multi-Layer Perceptron(Sa Res MLP)network,which employs attention mechanisms to mine important patches in medical images to better explain which parts determine image decisions.Specifically,it embeds the attention mechanism into the model to mine the important image patches and connects the attention mechanism with the loss function using a loss-based attention mechanism to reduce the inconsistency between interpretability and classification tasks.Experiments on several natural image and medical image datasets demonstrate the effectiveness and interpretability of the proposed method.(3)A medical image classification assisted diagnosis system is designed.Based on the research results of the above medical image classification model,this thesis further designs a visualized assisted classification platform,which mainly includes two modules of data enhancement and image classification to provide the relevant personnel with simple and convenient assisted diagnosis results and reduce the rate of misdiagnosis and missed diagnosis.