Non-invasive multi-phase CT artificial intelligence for predicting pre-treatment enlarged lymph node status in colorectal cancer: a prospective validation study.
Authors
Affiliations (10)
Affiliations (10)
- Department of General Surgery, Peking University Third Hospital, Beijing, China.
- Cancer Center, Peking University Third Hospital, Beijing, China.
- Beijing Key Laboratory for Interdisciplinary Research in Gastrointestinal Oncology (BLGO), Beijing, China.
- Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan, China.
- Department of General Surgery, Peking University Third Hospital, Beijing, China. [email protected].
- Cancer Center, Peking University Third Hospital, Beijing, China. [email protected].
- Beijing Key Laboratory for Interdisciplinary Research in Gastrointestinal Oncology (BLGO), Beijing, China. [email protected].
- Department of General Surgery, Peking University Third Hospital, Beijing, China. [email protected].
- Cancer Center, Peking University Third Hospital, Beijing, China. [email protected].
- Beijing Key Laboratory for Interdisciplinary Research in Gastrointestinal Oncology (BLGO), Beijing, China. [email protected].
Abstract
Benign lymph node enlargement can mislead surgeons into overstaging colorectal cancer (CRC), causing unnecessarily extended lymphadenectomy. This study aimed to develop and validate a machine learning (ML) classifier utilizing multi-phase CT (MPCT) radiomics for accurate evaluation of the pre-treatment status of enlarged tumor-draining lymph nodes (TDLNs; defined as long-axis diameter ≥ 10 mm). This study included 430 pathologically confirmed CRC patients who underwent radical resection, stratified into a development cohort (n = 319; January 2015-December 2019, retrospectively enrolled) and test cohort (n = 111; January 2020-May 2023, prospectively enrolled). Radiomics features were extracted from multi-regional lesions (tumor and enlarged TDLNs) on MPCT. Following rigorous feature selection, optimal features were employed to train multiple ML classifiers. The top-performing classifier based on area under receiver operating characteristic curves (AUROCs) was validated. Ultimately, 15 classifiers based on features from multi-regional lesions were constructed (Tumor<sub>N, A</sub>, <sub>V</sub>; Ln<sub>N</sub>, <sub>A</sub>, <sub>V</sub>; Ln, lymph node; <sub>N</sub>, non-contrast phase; <sub>A</sub>, arterial phase; <sub>V</sub>, venous phase). Among all classifiers, the enlarged TDLNs fusion MPCT classifier (Ln<sub>NAV</sub>) demonstrated the highest predictive efficacy, with AUROCs and AUPRCs of 0.820 and 0.883, respectively. When pre-treatment clinical variables were integrated (Clinical_Ln<sub>NAV</sub>), the model's efficacy improved, with AUROCs of 0.839, AUPRCs of 0.903, accuracy of 76.6%, sensitivity of 67.7%, and specificity of 89.1%. The classifier Clinical_Ln<sub>NAV</sub> demonstrated well performance in evaluating pre-treatment status of enlarged TDLNs. This tool may support clinicians in developing individualized treatment plans for CRC patients, helping to avoid inappropriate treatment. Question There are currently no effective non-invasive tools to assess the status of enlarged tumor-draining lymph nodes in colorectal cancer prior to treatment. Findings Pre-treatment multi-phase CT radiomics, combined with clinical variables, effectively assessed the status of enlarged tumor-draining lymph nodes, achieving a specificity of 89.1%. Clinical relevance statement The multi-phase CT-based classifier may assist clinicians in developing individualized treatment plans for colorectal cancer patients, potentially helping to avoid inappropriate preoperative adjuvant therapy and unnecessary extended lymphadenectomy.