BI-RADS Classification Of Breast MRI Based On Deep Learning

With breast cancer now leading the number of new cases of all cancers,early diagnosis has become especially important.Magnetic resonance imaging(MRI)is one of the commonly used techniques for breast cancer diagnosis,but it requires specialized physicians to perform readings and other tasks,exacerbating the strain on medical resources.The development of deep learning technology has opened up new possibilities for computer-aided diagnosis,which is expected to improve diagnostic efficiency and reduce the workload of physicians.However,most of the existing deep learning-based tumor classification tasks are binary tasks,and coarse-grained classification methods cannot provide enough fine-grained information to help doctors make more accurate diagnoses.In addition,modern neural networks suffer from insufficient confidence calibration to indicate when the model may be wrong,which may lead to misdiagnosis or missed diagnosis.To address the above issues,the contributions and innovations of the thesis are as follows:First,the thesis proposes a confidence-based calibration of the breast MRI BI-RADS classification model,which is guided by the BI-RADS evaluation criteria,generates finegrained pseudo-labels for guiding the training of the model by recording information about the malignancy likelihood of the training sample tumors generated during the training of the model,and introduces a single-peaked distribution to further constrain the classifier,which effectively improves the model’s ability to capture fine-grained information,and the leap from binary classification to ordered multi-classification is achieved.The experimental results show that the model can significantly improve the calibration of the model while maintaining a reasonable confidence distribution.Second,based on contrasting learning,the thesis proposes a homologous contrast algorithm to extend the contrast of a single image to the contrast between all images of the same case,as a way to improve the ability of the model to learn better feature representation with a smaller amount of data.Meanwhile,based on the BYOL contrast framework,the thesis builds a homologous contrast classification model for breast MRI,and experiments based on binary classification and BI-RADS classification demonstrate the effectiveness of the homologous contrast algorithm.Third,due to the special characteristics of breast MRI,the research results and conclusions of general image classification cannot be directly applied to this task.The thesis conducted comparison experiments on multiple image enhancement methods based on breast MRI;the thesis designed an attention-based feature fusion algorithm and a recurrent neural network-based feature fusion algorithm and experimentally explored the role of multiple feature fusion methods.Based on the observation of breast MRI,a simple automatic segmentation algorithm is proposed in this thesis,which can automatically remove the tissues in the chest cavity in the image,reduce their interference in the model recognition of tumors,and effectively improve the stability of classification.