A multi-class classification algorithm based on hematoxylin-eosin staining for neoadjuvant therapy in rectal cancer: a retrospective study

PeerJ

article

Yihan Wu¹  Xiaohua Liu³  Fang Liu⁴  Yi Li²  Xiaomin Xiong²  Hao Sun⁵  Bo Lin²  Yu Li⁴  Bo Xu¹ 

[1] School of Medicine, Chongqing University;Chongqing Key Laboratory of Intelligent Oncology for Breast Cancer, Chongqing University Cancer Hospital;Bioengineering College of Chongqing University;Department of Pathology, Chongqing University Cancer Hospital, School of Medicine, Chongqing University;Department of Gastrointestinal Cancer Center, Chongqing University Cancer Hospital

关键词: Rectal cancer;    Neoadjuvant therapy;    Deep learning;    Multi-class classification algorithms;    Pathology;   

DOI  :  10.7717/peerj.15408

学科分类：社会科学、人文和艺术（综合）

来源: Inra

【 摘 要 】

Neoadjuvant therapy (NAT) is a major treatment option for locally advanced rectal cancer. With recent advancement of machine/deep learning algorithms, predicting the treatment response of NAT has become possible using radiological and/or pathological images. However, programs reported thus far are limited to binary classifications, and they can only distinguish the pathological complete response (pCR). In the clinical setting, the pathological NAT responses are classified as four classes: (TRG0-3), with 0 as pCR, 1 as moderate response, 2 as minimal response and 3 as poor response. Therefore, the actual clinical need for risk stratification remains unmet. By using ResNet (Residual Neural Network), we developed a multi-class classifier based on Hematoxylin-Eosin (HE) images to divide the response to three groups (TRG0, TRG1/2, and TRG3). Overall, the model achieved the AUC 0.97 at 40× magnification and AUC 0.89 at 10× magnification. For TRG0, the model under 40× magnification achieved a precision of 0.67, a sensitivity of 0.67, and a specificity of 0.95. For TRG1/2, a precision of 0.92, a sensitivity of 0.86, and a specificity of 0.89 were achieved. For TRG3, the model obtained a precision of 0.71, a sensitivity of 0.83, and a specificity of 0.88. To find the relationship between the treatment response and pathological images, we constructed a visual heat map of tiles using Class Activation Mapping (CAM). Notably, we found that tumor nuclei and tumor-infiltrating lymphocytes appeared to be potential features of the algorithm. Taken together, this multi-class classifier represents the first of its kind to predict different NAT responses in rectal cancer.

【 授权许可】

CC BY