Interpretative Classification Of Pathological Images Based On Deep Learning

Real-time,objective,accurate diagnostic results,and providing interpretable features for the results help doctors to take symptomatic treatment.At present,pathological examination is still one of the important methods for the diagnosis of colonic lesions.However,the existing detection methods not only take a long time,but also rely heavily on the subjective interpretation of pathologists,leading to misdiagnosis and missed diagnosis.To address these issues,a residual neural network combined with the attention mechanism was put forward to predict the labels of eight kinds of colonic pathology images.Meanwhile,Gradient-weighted Class Activation Mapping algorithm is used to predict the labels of eight kinds of colonic pathology images.The Grad-CAM generated heat maps of attention can also decode the patterns learned by the network into interpretable category information.Experiments show that the network reaches an accuracy98.78% and an AUC value of 0.9983,respectively,for eight kinds of colon tissue "image blocks" including colon cancer,which proved the effectiveness of the model in the multi-classification of pathological images.Meanwhile,the generated attention heat map was basically consistent with the determination method of pathologists.It provides interpretability for the classification results of pathological images.It is difficult for colorectal cancer identification to realize both high precision prediction and explainable reasoning,which is of great significance in clinical medicine.However,post-hoc interpretability analysis of the model using attention heatmaps fails to make category-specific feature extraction,nor to explain the internal decision-making process of the network.The network based on prototypical decision trees can not only provide a classification basis for the model by extracting category features,but also the prototypical inference process can provide an internal decision process for deep learning models.To this end,this paper proposes the Ensemble Adaptive Boosting Prototype Tree(En ABPT)framework,which can predict the labels of eight types of colon pathology image blocks by integrating five serial neural prototype trees,and by visualizing the decision process of each base learner.Provide interpretable reasoning processes for the model.The results showed that the specific accuracy and AUC value of the method for eight types of colorectal pathological image blocks reached 97.8% and0.9668,respectively.The problem of "accuracy-interpretability trade-off" is addressed.The superiority of this framework can provide a new paradigm for interpretable reasoning and high-precision prediction of pathological image patches in pathology.