Application Of Deep Learning-based Radiomics In Tumor Classification

The past few decades have seen revolutionary advances in medicine and healthcare.During this time,the actual causes behind many diseases were revealed,new diagnostics were devised,and new drugs were developed.Even with all these achievements,cancer still plagues us and our health remains vulnerable to them.With the continuous development of hardware and software technology,medical imaging technology has become an important means of cancer diagnosis and treatment.Commonly used medical imaging includes X-ray detection,computed emission tomography（CT）imaging,magnetic resonance imaging（Magnetic Resonance Imaging,MRI）,ultrasound imaging,histopathological testing,etc.Medical image analysis technology based on these images has become an indispensable technical means in medical research,clinical disease diagnosis,and treatment.Among them,radiomics combines medical image analysis and machine learning technology to extract highthroughput meaningful information from medical data and conduct corresponding analysis and has achieved many results in clinical experiments.At present,radiomics based on traditional machine learning uses manually extracted image features,while with the development of deep learning,automated feature extraction and efficient endto-end prediction capabilities have been successfully applied in medical image analysis.Deep learning radiomics methods are of great significance to the development of computer-aided diagnosis systems.The application of deep learning-based radiomics technology in the auxiliary diagnosis of cancer is the research topic of this paper.According to the characteristics of small samples and high specificity of tumor images,it is the content of this paper to apply or propose new radiomics and deep learning methods and to analyze the medical images of glioma and colorectal cancer.as follows:1.Aiming at the problem of the difficulties of the classification of glioma subtypes,two convolutional neural network models were trained based on multimodal magnetic resonance images and histopathological images,and the performance of the models was improved through ensemble learning,and finally,an end-to-end model was obtained.The peer-to-peer glioma subtype classification prediction model combined with pathological images won first place in the 2019 Precision Medicine Classification Competition,which can provide effective assistance for the clinical glioma subtype classification.2.Aiming at the prediction of MGMT methylation and molecular typing of 1p/19 q co-deletion in glioma on multimodal magnetic resonance images,a Dense Net network based model fused with attention mechanism was proposed.Deep learning based radiomics methods achieve the AUCs of 0.73 and 0.87 on MGMT methylation and1p/19 q co-deletion datasets,respectively,an average improvement over baseline radiomics models of 21%,which is comparable to the performance of the current SOTA method.3.To build a prediction model for the binary classification of the T staging of colorectal cancer,combined with the imaging characteristics of colorectal cancer,a residual network model combining multi-scale and non-regional modules was proposed,using pyramid pooling（Spatial Pyramid Pooling,SPP）extracts the multi-scale features of the image,and adds a non-local module to improve the feature extraction efficiency of the network.The optimal model achieved an AUC of 0.84,and the prediction effect of the model’s binary T staging was comparable to that of senior doctors in the imaging department.