Predicting lymph node metastasis in colorectal cancer using case-level multiple instance learning.

Area of Science:

• Digital pathology
• Artificial intelligence in oncology
• Colorectal cancer research

Background:

• Accurate lymph node metastasis (LNM) prediction is vital for locally advanced (T3/T4) colorectal cancer (CRC) management.
• Traditional histopathology and slide-level deep learning struggle with sparse, critical metastatic features.

Purpose of the Study:

• Develop and validate a case-level multiple-instance learning (MIL) framework.
• Mimic pathologist comprehensive review for improved T3/T4 CRC LNM prediction.

Main Methods:

• Retrospective analysis of whole-slide images from 130 T3/T4 CRC patients.
• Case-level MIL framework using CONCH v1.5 and UNI2-h deep learning models.
• Integration of pathological features with clinical data; performance evaluated by AUC.

Main Results:

• Case-level MIL framework outperformed slide-level training (CONCH v1.5 AUC: 0.899 vs 0.814).
• Integrating pathology and clinical data enhanced prediction (top model AUC: 0.904 vs clinical-only AUC: 0.584).
• Model-identified regions aligned with pathologist-confirmed high-risk histopathological features.

Conclusions:

• Case-level MIL framework offers a superior method for LNM prediction in advanced CRC.
• This approach shows potential for risk stratification and guiding therapy decisions.
• Further validation of the framework is warranted.