| Breast cancer,as a common malignant tumor in women,seriously endangers women’s life and health due to its high morbidity and mortality.The distinction between benign and malignant breast tumors is a necessary link in the diagnosis of breast cancer and the basis for formulating treatment plans.Breast cancer is divided into four molecular subtypes,and different subtypes have significant differences in prognostic effect and recurrence risk.However,the determination of molecular subtypes usually requires invasive pathological tests,and it is of great significance to develop noninvasive and three-dimensional methods to determine the molecular subtypes of tumors.Dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI),as a non-invasive examination method,has the advantages of being sensitive to lesions and reflecting tumor angiogenesis,and has good diagnostic value for breast cancer.Deep learning has been widely used in breast cancer diagnosis and prediction research in recent years,and achieved good results,but its automatic feature extraction mechanism is out of control,resulting in poor interpretability.The regional features of DCE-MRI tumors provide an important reference for imaging diagnosis.If the deep learning model pays more attention to this region,the interpretability of the model can be effectively improved.In this study,based on the deep feature interpretable method,under the guidance of tumor region information,a diagnostic model based on deep feature deconvolution mapping and semantic information guidance was designed to realize the prediction of benign and malignant breast tumors and molecular subtypes of breast cancer.The class activation map,as a model visualization method,shows the model’s attention region,and the similarity of this region to the tumor region reflects the correlation between the model decision and the tumor region information,verifying the interpretability of the model.The specific content of this study is as follows:(1)Research on breast DCE-MRI image segmentation method based on deep learningIn order to extract accurate tumor regions,a breast image tumor segmentation network based on contextual pyramid fusion is proposed,which does not need to pre-label tumor regions,and improves the noise and incomplete segmentation problems existing in traditional threshold segmentation methods.In addition,a tumor region expansion algorithm based on breadth-first search was designed to obtain different ranges of tumor-related regions for ablation experiments on the region and the performance of the diagnostic model.The experimental results show that the average Dice coefficient of the tumor segmentation network is 0.819,and the average Hausdorff distance(Hausdorff Distance,HD)is 4.404,which is better than the threshold segmentation method(Dice=0.602,HD=6.047),which improves the segmentation effect.(2)Research on intelligent diagnosis model based on deep feature deconvolution mapping and tumor region information guidanceThis study proposes a diagnostic model based on deep feature deconvolution mapping and tumor region information guidance.During the training process,the similarity loss between the tumor region and the result of deep feature deconvolution mapping is calculated to achieve model attention region guidance.The guided model has a macro average(Macro average)AUC of 0.961 in the three classification tasks of benign and malignant breast tumors and healthy breasts,which is better than the unguided model(Macro average AUC=0.951).In the Luminal A subtype binary classification task,the AUC reached 0.792,better than the unguided model(AUC=0.754),indicating that the model has better classification performance.The average Intersection over Union(Io U)of the high activation area of the model’s activation map and the tumor area is 0.114,which is higher than that of the unguided model(Io U=0.049).The model’s attention pays more attention to the information of the tumor area,reflecting more Good interpretability.(3)Research on intelligent diagnosis model guided by semantic information and tumor region informationThis study proposes a diagnostic model based on semantic information and tumor region information guidance,using the deep feature mapping results of fusion prediction category semantic information and tumor region calculation similarity loss,which strengthens the guiding role of tumor region information compared with the deconvolution mapping method.The macro average(Macro average)AUC in the three classification tasks of benign and malignant breast tumors and healthy breasts is 0.967,which is 0.015 higher than that of the unguided model.In the Luminal A subtype binary classification task,the AUC reached 0.834,which is 0.080 higher than that of the unguided model.The average Io U between the high activation region and the tumor region of the model’s activation map is 0.353,which is significantly higher than that of the unguided model(Pvalue=3.597×10-20).Through ablation experiments,different expansion ratios of the tumor area were set to evaluate the model performance.When the Expansion ratio is greater than 50%,the classification performance of the model gradually decreases.The best Expansion ratio is 20%,and the corresponding AUC is 0.837.The experimental results show that the model further improves the classification performance and interpretability,and setting too large tumor-related region-guided training will reduce the classification performance of the model.This study proposes a deep interpretable tumor diagnosis model,which is applied to the prediction of benign and malignant breast tumors and breast cancer molecular subtypes,and the interpretability of the model is reflected according to the correlation between the model attention and the tumor region.The experimental results show that the guidance of tumor region information improves the classification performance and interpretability of the model,making the diagnosis and prediction results more reliable,which is conducive to the further promotion of deep learning in breast cancer diagnosis research. |
PDF Full Text Request