Deep learning based on MRI for assessing the prognostic value of lateral lymph nodes in rectal cancer

Guanzhong Qiao^1*Lili Feng²Zhen-Hui Li³Qiong Wu¹Yulin Liu⁴Jie Zhao¹Hao Jiang¹Ke Zhao⁵Yanfen Cui⁶Huijie Jiang^1*

• ¹The Second Affiliated Hospital of Harbin Medical University, Harbin, China
• ²Sun Yat-sen University Cancer Center, Guangzhou, China
• ³Yunnan Cancer Hospital, Kunming, China
• ⁴Sun Yat-sen University Sixth Affiliated Hospital, Guangzhou, China
• ⁵Guangdong Provincial People's Hospital, Guangzhou, China
• ⁶Shanxi Cancer Hospital, Taiyuan, China

Objectives: Accurate preoperative evaluation of positive lateral lymph node (LLN) is crucial for optimizing treatment strategies in rectal cancer. Traditional methods, such as MRI T2-weighted imaging (T2WI), face limitations like interobserver variability and difficulty detecting small or occult metastases. Deep learning (DL) may provide a more efficient and precise alternative. Methods: In this multicenter, retrospective study, images from 1,000 patients across five centers were annotated to train a DL model for identifying and segmenting LLN. The model was tested on images from 480 patients in a validation cohort. Kaplan-Meier analysis compared disease-free survival (DFS) and overall survival (OS) between LLN-positive and LLN-negative groups, while Cox regression identified prognostic factors for DFS and OS. Results: The DL model achieved an accuracy of 87.5% and a specificity of 73.8% in predicting LLN positivity, demonstrating high diagnostic performance. Both univariate and multivariate Cox regression analyses identified LLN status, circumferential resection margin (CRM), and tumor downstaging (TD) as independent prognostic factors. Kaplan-Meier analysis showed patients with positive LLNs had worse outcomes, with 3-year DFS of 57.66% vs. 81.66%, and 5-year OS of 61.62% vs. 84.82% compared to LLN-negative patients. Conclusions: The DL model effectively predicts positive LLNs, offering an efficient alternative to traditional methods and supporting preoperative decision-making. Its clinical implementation could enhance risk stratification and personalize therapeutic strategies for rectal cancer patients.