| Prostate cancer is the second most common cancer type in men,and Gleason grading is a standard prognostic factor and also helpful for recommending a treatment regime.In the current clinical setting,the Gleason score is based on manual grading of gland differentiation in histopathological specimens under a microscope by trained pathologists.However,this process is time-consuming and highly subjective,which is prone to misdiagnosis and missed diagnosis.Therefore,it is of great significance to develop a quantitative,accurate,and reproducible Gleason grading model for prostate cancer to assist pathologists in diagnosis.The thesis consists of two aspects.First,the important domain knowledge related to gland structure is introduced into the auxiliary diagnosis model,then a Gleason grading method for prostate cancer based on semantic and structural features is proposed.Second,to deal with the domain shift caused by tissue preparation and the scanning process,a sample-adaptive Gleason grading method for prostate cancer is proposed.In the first work,this thesis constructs a dual-branch network including a multi-scale semantic feature extractor and a multi-level structural feature extractor.The former is used to extract high-level semantic features,and the latter utilizes graph convolution to extract hierarchical structure features from constructed cell graphs.Features are refined by the channel attention mechanism for the Gleason grading task.Through the above improvements,the macro F1 and micro F1 on the test set reach 0.7365 and 0.7792,respectively,which are superior to several deep learning-based methods for Gleason grading.In the second work,this thesis combines domain-related content information,style information,and topological information of cells,then introduces a total variational regularization term to train a cross-domain image generation model.The content and style information is calculated by a pre-trained VGG16 model,while the cell topology information is calculated by a nucleus segmentation network.Using this model,a large number of labeled images in the source domain can be automatically converted to the target domain,and the classification accuracy of the test samples in the target domain can be significantly improved.Comparative experiments show that this scheme outperforms widely used staining normalization methods as well as advanced adversarial training methods,reaching 0.6055 and0.7650 for macro F1 and micro F1,respectively,on the target domain test set.The two aspects of this work have effectively improved the performance of the automatic Gleason grading system for prostate cancer,which has important clinical significance.In addition,related methods have certain generality and have the potential to be extended to the development of computational pathology models for other diseases. |
