A novel artificial intelligence approach to the prediction of lymph node metastasis using whole-slide imaging in patients with T1 colorectal cancer.

INTRODUCTION

The accurate assessment of lymph node metastasis (LNM) in T1 colorectal cancer (CRC) is critical to guide any surgery required following endoscopic resection. However, pathology-based risk stratification is subject to interobserver variability. Therefore, we aimed to develop and validate a stacking-based artificial intelligence (AI) model that integrates the results of the analysis of hematoxylin and eosin (HE)-stained whole-slide images (WSIs) with clinical features.

METHODS

Patients with T1 CRC who had their tumors resected at Showa Medical University Northern Yokohama Hospital between 2001 and 2018 were used to train the model. Internal validation was performed using data from consecutive patients collected during 2018-2021, and external validation was performed using data collected at two regional hospitals during 2018-2023. The model used the Multiple Instance Self-Training (MIST) score, sex, and tumor location (colon vs. rectum) as inputs, and logistic regression, XGBoost, and random forest classifiers as analyses.

RESULTS

A total of 593 patients were used for training. LNM was present in 15% (15/100) of the internal cohort and 8.0% (2/25) of the external cohort. The stacking model generated areas under the curve (AUCs) of 0.68 (internal) and 0.80 (external), which outperformed guideline-based stratification (AUCs of 0.52 and 0.52, respectively).

CONCLUSIONS

The AI model, which integrates WSIs and clinical data, is an accurate, objective means of LNM risk prediction for patients with T1 CRC. It may complement pathology-based assessments and reduce overtreatment. However, further validation through prospective multicenter studies is warranted.

REGISTRATION

The University Hospital Medical Network Clinical Trials Registry (UMIN 000046992).