An Artificial Intelligence System to Predict Pathologic Response Using Longitudinal MRI in Locally Advanced Rectal Cancer: A Multicenter Study

Background: After neoadjuvant chemoradiotherapy (NCRT) for locally advanced rectal cancer (LARC), about 20% of patients could achieve pathologic complete response (pCR) without residual tumor. This led to the proposition of organ preservation by omitting subsequent radical surgery. However, the clinical implementation of this strategy is significantly hindered by the lack of tools to reliably identify which patients would achieve pCR before surgery. The aim of this study was to develop and validate an artificial intelligence (AI) system for accurate prediction of pathologic response in patients with LARC.    Methods: We conducted a retrospective, multi-cohort study of 1201 patients recruited from four hospitals in China. First, based on 112596 images from 638 patients from the 6th Hospital of Sun Yat-sen University, Guangzhou (SYSU6H), we trained a multi-task deep learning model that allows simultaneous tumor segmentation and response prediction using longitudinal multi-parametric MRI (T1 weighted with and without contrast, T1 contrast-enhanced, T2, DWI) acquired before and after NCRT. The model’s diagnostic performance was then independently tested in an internal validation set from SYSU6H (32912 images from 186 patients) and three external validation sets [the Cancer Center of Sun Yat-sen university, Guangzhou (SYSUCC), 41114 images from 235 patients; the Nanfang Hospital, Guangzhou (NFH), 13668 images from 79 patients; and the 1st Affiliated Hospital of Soochow University, Suzhou (SDFYY), 11932 images of 63 patients]. The model was compared with two experienced radiologists in 265 randomly selected patients from three validation sets. Further, the model’s performance for predicting pathologic downstage and non-response, as well as its prognostic significance were evaluated. Findings: The deep learning model achieved high performance in predicting pCR, with an area under the curve (AUC) values of 0·969 (0·942~0·996) for the SYSU6H internal validation set, and 0·946 (0·942~0·996), 0·943 (0·888~0·998), 0·919 (0·840~0·997) for the SYSUCC, NFH and SDFYY external validation sets, respectively. The model performed equally well in subgroups of patients defined by receipt of radiotherapy, clinical response, tumor location and age. The model also showed superior performance to radiologists and can be used to improve human reading. The continuous deep learning score was highly predictive of the degree of pathologic response, and was prognostic of survival independent of clinicopathologic factors.   Interpretation: The AI system is accurate in identifying patients with pCR after NCRT. Its potential clinical utility of selecting patient for organ preservation warrants further investigation in future randomized trials. Funding Information: This research was supported by the National Key R&D Program of China (2017YFC1308800), National Natural Science Foundation of China（81970452, 81972212, 81770656). Declaration of Interests: We have no declarations of interest to declare. Ethics Approval Statement: All patients completed a written informed consent form before entering the study. The study was approved by the local medical ethics committee of the four centers and was conducted in accordance to the Declaration of Helsinki and good clinical practice.