Image-based AI has thrived as a potential revolutionised tool for predicting the status of molecular biomarkers, which effectively helps group patients with targeted medical treatments. However, methods based on haematoxylin and eosin-stained (H&E) whole-slide images (WSIs), usually accommodate the entire images directly as model inputs, which could be inefficient as many image patches are not informative or even irrelevant to the biomarkers. In this study, we introduced a new concept of region of biomarker-relevance (ROB) to represent the morphological regions most related to a specific biomarker and to serve as a fundamental element for WSI-based biomarker prediction. We embodied the ROB theory with framework titled saliency ROB search (SRS) to effectively predict biomarker status. We exemplified the SRS by analyzing various adenocarcinoma (LUAD) biomarkers, and achieved better performances than the-state-art of AI approaches. This indicates that on the basis of ROB, AI can reach a better molecular biomarker prediction precision from pathological images.
Keywords: molecular biomarker prediction, lung adenocarcinoma, computional pathology, deep learning, whole slide image, subtype lesion detection, region of biomarker-relevance selection, adaptive
Gan, Jiefeng and Wang, Hanchen and Yu, Hui and He, Zitong and Zhang, Wenjuan and Ma, Ke and Zhu, Lianghui and Bai, Yutong and Zhou, Zongwei and Yuille, Alan L. and Bai, Xiang and Wang, Mingwei and Yang, Dehua and Chen, Yanyan and Chen, Guoan and Lasenby, Joan and Cheng, Chao and Wu, Jia and Zhang, Jianjun and Wang, Xinggang and Chen, Yaobing and Wang, Guoping and Xia, Tian, Focalizing Regions of Biomarker-Relevance Facilitates Biomarker Prediction on Histopathological Images. Available at SSRN: https://ssrn.com/abstract=4361627 or http://dx.doi.org/10.2139/ssrn.4361627
This version of the paper has not been formally peer reviewed.
Download This Paper