| Pathological diagnosis is considered the "gold standard" for cancer diagnosis,playing a crucial role in determining patient treatment and prognosis.The traditional method of manual slide examination requires pathologists to observe under a microscope,which is time-consuming,labor-intensive,and subject to the subjective factors of the doctors,affecting the accuracy.With the emergence of whole slide images(WSIs)and the development of computer-aided diagnosis(CAD)techniques,pathologists can now utilize technologies such as artificial intelligence to assist in their diagnosis,enabling them to analyze images while sitting in front of a computer.This approach greatly enhances the efficiency and accuracy of cancer diagnosis.WSIs pose a significant challenge to traditional models due to their extremely high resolution.Therefore,the usual approach is to divide them into smaller blocks for processing.In this process,it is crucial to train a high-accuracy model for small block-level pathological image classification.The morphology and texture features of cancerous tissues in WSI are complex,and existing models often cannot fully express these features,resulting in limited identification accuracy.In addition,WSI annotation is a difficult process,and inaccurate labeling can result in a significantly increased likelihood of assigning incorrect labels to small blocks,severely affecting the training of small block-level pathological image classification models.Therefore,this paper aims to improve the accuracy of identifying cancerous regions in WSI starting from the problem of small block-level pathological image classification.The main research contents are as follows:(1)To address the problem that the deep layer of existing deep learning models fails on small block-level pathological images,the Frequency-domain Attention Network(FA-Net)is proposed to introduce the frequency-domain information of images through a branch that extracts frequency-domain features,which will guide the feature expression of images in the spatial domain in the form of an attention mechanism that will Some features that are originally difficult to express are expressed explicitly in the spatial domain to improve the overall feature expression capability of the model.Experiments on two datasets,PCAM and NMIWSI,show that FA-Net has stronger feature expression ability than other models for pathological images.(2)For addressing the issue of mislabeling arising from dividing fullfield pathological images into small blocks,a self-supervised labeling tolerance mechanism based on class activation is proposed,inspired by class activation mapping.The proposed mechanism utilizes the classification probabilities obtained by the model as soft labels for additional supervision,reducing the risk caused by mislabeling.Furthermore,a gradient class activation-based model self-supervised labeling fault tolerance mechanism is introduced to derive classification probabilities using gradient information.Unlike other methods that generate soft labels,the proposed model self-supervised labeling fault tolerance mechanism entirely utilizes the model’s information for generating soft labels without introducing additional model or data domains.Experiments on two datasets,PCAM and CIFAR-10 N,demonstrate that these two label tolerance mechanisms not only alleviate the effect of mislabeling in the pathological image dataset but also have the same effect on the noisy labeled dataset of natural images.(3)The proposed methods and models are applied to the CAMELYON16 dataset to identify cancerous regions in WSI.Firstly,small block-level pathological images are sampled from the WSI to train the corresponding models.Then,the trained models are used to predict the divided small block-level pathological images in the test set,and finally,the predicted results of small blocks are combined into a heatmap of cancerous regions.The experiments show that the proposed methods and models significantly improve the recognition performance compared to the baseline models. |
PDF Full Text Request