Automatic Assessment Of Cancer Malignancy Based On Pathological Image Analysis

Adenocarcinoma is a type of cancer that often occurs in epithelial gland tissues.It is common in malignant tumors such as colorectal cancer,prostate cancer,breast cancer,and lung cancer.In terms of pathological pattern,the occurrence of adenocarcinoma is often accompanied by poor or undifferentiated gland structure.The glands are abnormally stretched and distorted,and the cavity structures are invaded by the nuclei,causing it to shrink or even disappear.This situation is directly related to the degree of malignancy of adenocarcinoma.Therefore,the degree of differentiation of glandular structure is a decisive factor for pathologists to determine the grade of adenocarcinoma and decide the treatment plan in clinical pathological diagnosis.In this paper,to realize the automated grading of the malignant degree of adenocarcinoma,the study evolves from the degree of gland structure differentiation which is a pathological diagnostic criteria.On the one hand,this paper designs domain-specific hand-crafted features and proposes an automated gleason grading method for prostate adenocarcinoma based on the statistical representation of homology profile.This method first calculates the homology components of pathological images based on the homology profile algorithm,and quantifies the degree of gland structure differentiation by describing the topological arrangement of nuclei around the gland.Then it employs statistical methods for secondary characterization on the obtained homology sequence characteristics.The method finally models the weighted Knearest neighbor classifier algorithm for automated gleason grading of pathological images on adenocarcinoma.Experimental results show that this method performs better in grading and features robustness,compared with unsupervised learning method SSAE,supervised method DLGg and traditional pathomics-based method MATF.In addition,the proposed homology feature representation method has a strong biological basis and is highly interpretable.On the other hand,inspired by the idea of accurately describing the automatic segmentation of glandular structures for diagnosis,this paper proposes an automated assessment method for colorectal adenocarcinoma based on multi-task learning and prior knowledge.Firstly,this method employs the backbone network for feature extraction,and the extracted features are respectively sent to the segmentation branch and classification branch for automated gland structure segmentation and grading of adenocarcinoma images.Meanwhile,the automated gland structure prediction from the segmentation branch acts as the prior knowledge,and it is encoded as spatial attention and merged into the classification branch to constrain the reasoning.The experimental results verify the effectiveness of the prior knowledge-aware multi-task network,achieving the highest accuracy of 97.04% and AUC value of 0.9971 on the test set.Moreover,the idea of prior knowledge constraint in this paper is also interpretable.The automated assessment grading methods based on pathological image analysis are proposed for cancer malignancy.Importantly,the glandular structure differentiation acting as the pathological diagnosis criteria is fundamental to them.The methods thus can provide pathologists with interpretable auxiliary diagnosis support.