Weakly Supervised Learning Algorithm And Its Application On Histopathology Images

Histopathological examination is known as the gold standard for cancer diagnosis.The use of deep learning techniques to accelerate the diagnosis of digital pathology slides has significant value for various clinical applications.This paper mainly studies weakly supervised learning and its application in pathological sections.By analyzing current weakly supervised learning algorithms in the field of pathology,it is found that although the current two-stage weakly supervised neural network algorithm can effectively speed up the training process,there are still the following problems:(1)The pre-trained model cannot adapt to the distribution of pathological images.At present,two-stage networks often use model weights trained on natural images and cannot adapt to the characteristics of pathological images.(2)The uneven distribution of label information in pathological sections at both global and local levels.The unique segmentation mechanism of the twostage network also makes it difficult for label information to guide the generation of image representation by the first-stage network.Based on the above analysis,this thesis studies the data preprocessing process of pathological sections,the pathological section pretraining model based on neural networks and the self-cleaning weakly supervised learning model.Finally,a glioma grading system is designed and implemented.The main work and contributions are summarized as follows:1.The preprocessing of digital pathology slides has been achieved.Starting from publicly available datasets and datasets organized by hospitals,the processes of data cleaning,image segmentation,image partitioning,and image feature extraction for digital pathology slides have been implemented.The results of the preprocessing of digital pathology slides are verified through image visualization methods.Through the preprocessing process,the storage space occupied by pathological sections is greatly reduced.The Camelyon dataset is reduced from its original 544 GB to 3.4GB.2.To address the issue of pre-trained models being unable to adapt to the distribution of pathological images,a pre-trained model for pathological sections of deep neural networks has been implemented.Contrastive learning is used for self-supervised learning on unsupervised image blocks.A loss function based on Chebyshev distance and a color discrimination loss function are proposed.This compensates for the performance loss caused by color drift during the staining process of digital pathology slides and also compensates for the loss of spatial relative position information during the data preprocessing process.In experimental verification,the proposed model’s AUC increased from 0.7 to0.81 compared to the baseline model,and its best,worst,and average values in 10-fold cross-validation all exceeded those of mainstream self-supervised learning models by 1%,demonstrating robust performance.3.To address the issue of uneven distribution of label information in pathological sections at both global and local levels,a self-cleaning weakly supervised learning model has been implemented.A method for identifying samples to be cleaned based on an asymmetric parallel network was proposed,combined with repeated pre-training of self-supervised learning models to improve the quality of image representation for difficult samples.The proposed model resulted in a 3% and 2% improvement for methods such as BYOL and Simsiam,respectively,while also bringing the classification performance under linear classification probes closer to that of complex classification methods.4.A glioma automatic grading system is designed and implemented.The function of uploading,preprocessing and automatic classification of tumor sections is realized.