Multi-Size Deep Learning Based Preoperative Computed Tomography Signature for Prognosis Prediction of Colorectal Cancer

• DOI: 10.3389/fgene.2022.880093
• Publication Date: 2022-05-12T08:30:30Z
• Authors: Cheng-Hang Li; Du Cai; Min-Er Zhong; Min-Yi Lv; Ze-Ping Huang; Qiqi Zhu; Chuling Hu; Haoning Qi; Xiaojian Wu; Feng Gao
• Keywords: Nomogram

Background: Preoperative and postoperative evaluation of colorectal cancer (CRC) patients is crucial for subsequent treatment guidance. Our study aims to provide a timely rapid assessment the prognosis CRC with deep learning according non-invasive preoperative computed tomography (CT) explore underlying biological explanations. Methods: A total 808 CT (development cohort: n = 426, validation 382) were enrolled in our study. We proposed novel end-to-end Multi-Size Convolutional Neural Network (MSCNN) predict risk recurrence images (CT signature). The prognostic performance signature was evaluated by Kaplan-Meier curve. An integrated nomogram constructed improve clinical utility combining other clinicopathologic factors. Further visualization correlation analysis features paired gene expression profiles performed reveal molecular characteristics tumors learned MSCNN radiographic imaging. Results: showed that significant factor disease-free survival (DFS) prediction [development hazard ratio (HR): 50.7, 95% CI: 28.4-90.6, p < 0.001; HR: 2.04, 1.44-2.89, 0.001]. Multivariable confirmed independence value 30.7, 19.8-69.3, 1.83, 1.19-2.83, 0.006). Dimension reduction demonstrated high patients. Functional pathway further indicated presented down-regulation several immunology pathways. Correlation found mainly associated activation metabolic proliferative Conclusions: based can effectively Integration multi-omic data revealed some tumor be captured images.