Research On Pathological Image Classification Model Combined With Fine-grained Feature Learning

In the current medical clinical research on malignant diseases,pathological examination is the most accurate "gold standard" for cancer identification and diagnosis.It can provide celllevel diagnostic results.Therefore,medical pathological images are very important.With the emergence of full-slice digital images and the development of artificial intelligence technology,it is the trend of the times to combine deep learning with medical images.The accumulation of a large number of pathological images provides the basis for the application of deep learning algorithms.At the same time,deep learning can extract hidden disease diagnosis features from digitally managed medical images.The use of deep learning methods to assist medical pathological image analysis and diagnosis is of great clinical significance.Histopathological images have the characteristics of high resolution,large size,tight arrangement,bright colors,different shapes,texture details and rich information.However,the deep learning models used in the current medical pathological image analysis have limited ability to extract these pathological features.Aiming at the deficiencies of the current research status,this paper studies the improvement of the deep learning model in the analysis of medical pathology images based on the pathological images of cholangiocarcinoma.In this study,the model is improved by adding conditional random fields,feature pyramids,and fine-grained feature learning to conduct experiments,and a pathological image multiclassification model FGENet combined with fine-grained feature learning is proposed.The main content of this paper includes:1)Perform experiments on the traditional deep convolutional network,residual convolutional network,attention mechanism network,neural architecture search network,and analyze from the perspectives of accuracy,precision,recall,and F1-score. At the same time Refer to the network performance of each model to conduct a horizontal comparison of several networks,thereby selecting the model EfficientNet-b1 with the best overall performance as the benchmark model;2)Improve the model from the perspectives of adding conditional random fields to modeling spatial correlation,adding feature pyramids to introduce multi-scale information,combining fine-grained feature learning to suppress salient features,etc., and obtaining the best combined fine-grained features through performance comparison The learned pathological image multi-classification model FGENet;3)Optimize the model,use mixed-precision training to reduce memory usage,accelerate model convergence,and use Adam+SGD optimizer to further improve model training accuracy;4)Perform transfer learning on the TCGA liver cancer data set and Cameylon16 breast cancer data set,and compare and evaluate the results of the experiment,thus verifying that the classification model has a strong generalization ability.The experimental results show that the proposed pathological image multi-classification model FGENet combined with fine-grained feature learning achieves an average accuracy of 92.8% in the validation set.Among them,the accuracy of tumor tissue is 93.5%,the accuracy of liver tissue is 92.1%,and the accuracy of lymphatic tissue is 92.8%.The accuracy rate is 92.9%,the average precision rate is 93.2%,the average recall rate is 93.1%,and the average F1-score is 93.1%.Both TCGA liver cancer data and Cameylon16 breast cancer data sets are divided into two categories: tumor tissue and other tissues,and the average accuracy rates on the two data sets are 92.1% and 91.5%,respectively.The model proposed in this paper has good evaluation results on multiple data sets,indicating that the pathological image multi-classification model FGENet combined with fine-grained feature learning has strong pathological image feature extraction capabilities,which can improve the classification effect and share some ideas for further researches.