Preoperative Prediction of Lymph Node Metastasis in Colorectal Cancer with Deep Learning

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate preoperative diagnosis of lymph node metastasis (LNM) is crucial for colorectal cancer (CRC) treatment planning.
• Current methods using radiology imaging or genomic tests for LNM prediction are often unreliable or expensive.
• There is a need for accessible and accurate methods for predicting LNM in CRC patients.

Purpose of the Study:

• To develop an interpretable, multimodal artificial intelligence (AI) method for predicting lymph node metastasis (LNM) in colorectal cancer (CRC) patients.
• To integrate pathological image features with serum tumor-specific biomarkers for enhanced LNM prediction.
• To validate the AI model's performance across different medical centers.

Main Methods:

• A Multimodal Multiple Instance Learning (MMIL) model was developed to extract features from pathological images.
• Clinical biomarker data was jointly integrated with image-derived features for LNM prediction.
• The model was trained on a discovery cohort (1128 patients) and validated on an external cohort (210 patients).
• Model interpretability was achieved through heatmap generation.

Main Results:

• The MMIL model demonstrated high predictive accuracy, achieving AUCs ranging from 0.809 to 0.926 in the discovery cohort (T1-T4 CRC stages).
• The model showed good generalizability on the external validation cohort with AUCs ranging from 0.691 to 0.855 (T1-T4 CRC stages).
• The MMIL model outperformed traditional preoperative radiology-imaging diagnosis.

Conclusions:

• The developed MMIL model shows significant potential for early LNM diagnosis in CRC patients.
• Integrating pathological images and tumor-specific biomarkers offers an accessible and accurate approach for LNM prediction.
• The model's interpretability highlights its ability to identify key histomorphologic features relevant to LNM prediction.