基于多字典学习框架的肾透明细胞癌预后分析模型

Prognostic Analysis Model of Renal Clear Cell Carcinoma Based on Multi-Dictionary Learning

肾透明细胞癌是一种高度异质的肿瘤,具有复杂多变的临床表现。基于病理全切片图像的肾透明细胞癌自动预后分析,可辅助医生做出临床决策,从而达到更好的治疗目的。肾透明细胞癌的组织异构性使得针对预后分析任务的特征提取存在很大的挑战性。提出针对肾透明细胞癌病理全切片图像的多字典学习框架,自适应获取病理全切片图像的有效信息,进行肾透明细胞癌预后分析。该框架主要包括基于图像块水平的多字典学习和基于患者水平的生存模型构建两个阶段。利用癌症基因组图谱数据库的肾透明细胞癌数据集(TCGA-KIRC)中378例苏木素-伊红染色的全切片图像上进行评估,实验结果(C-index=0.681,AUC=0.751,P<0.05)优于现流行的各种生存模型,其中较传统的Boosted模型和随机生存森林模型,C-index指标分别提高0.138和0.155,AUC指标分别提高0.149和0.191;较Deep Surv和WSISA两个深度学习模型,C-index指标分别提高0.046和0.035,AUC指标分别提高0.096和0.090。所提出的方法可以更准确地对肾透明细胞癌进行预后分析。

Clear cell renal cell carcinoma(ccRCC)is a highly heterogeneous tumor with complex and variable clinical manifestations.Automatic histopathological whole slide image(WSI)analysis is a useful approach for pathologists to make diagnosis.However,feature extraction for the prognostic analysis of ccRCC is a challenging task due to the diversity of tissue structures in the histopathological images.In this work,a novel WSI-based multi-dictionaries learning framework was proposed to adaptively extract the effective features of WSI for prognostic analysis of ccRCC.This framework included multi-dictionaries learning stage based on patch level and survival model construction stage based on patient level.The proposed model was evaluated on 378 hematoxylin-eosin stained WSIs form Cancer Genome Atlas database(TCGA-KIRC).The C-index was 0.681,and AUC was 0.751(P<0.05).Compared with the traditional Boosted model and Random Survival Trees model,the improvements on C-index were respectively 0.138 and 0.155,and the improvements on AUC was respectively 0.149 and 0.191.Compared with the two deep learning model(DeepSurv and WSISA),the improvements on C-index were respectively 0.046 and 0.035,and the improvements on AUC was respectively0.096 and 0.090.The results showed that the proposed model achieved superior performance for prognostic analysis of renal clear cell carcinoma.