Multicenter Evaluation of a Weakly Supervised Deep Learning Model for Lymph Node Diagnosis in Rectal Cancer at MRI

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate lymph node (LN) staging is crucial for rectal cancer (RC) treatment planning.
• Preoperative MRI is essential for evaluating RC, but accurate LN diagnosis remains challenging.
• Existing methods often lack the precision needed for optimal patient management.

Purpose of the Study:

• To develop and validate the Weakly supervISed model DevelOpment fraMework (WISDOM) for lymph node diagnosis in rectal cancer.
• To assess the performance of the WISDOM model using preoperative MRI and postoperative pathologic data.
• To evaluate the incremental value of the WISDOM model in assisting radiologists with LN staging.

Main Methods:

• A retrospective study utilizing MRI (T2-weighted and diffusion-weighted imaging) and patient-level pathologic data from 1014 rectal cancer patients.
• Development of the WISDOM model incorporating weakly supervised learning techniques.
• Evaluation of binary and ternary N staging performance using area under the receiver operating characteristic curve (AUC) and concordance index (C index).

Main Results:

• The WISDOM model achieved an overall AUC of 0.81 and C index of 0.765.
• The model significantly outperformed junior radiologists (AUC=0.69, C index=0.689) and performed comparably to senior radiologists (AUC=0.79, C index=0.788).
• Assistance from the WISDOM model significantly improved both junior (AUC=0.80, C index=0.798) and senior radiologists' (AUC=0.88, C index=0.869) performance.

Conclusions:

• The WISDOM model demonstrates significant potential as an effective lymph node diagnosis tool for rectal cancer using routine MRI data.
• The model's ability to enhance radiologist performance highlights its practical clinical utility.
• WISDOM offers a promising approach to improve the accuracy of lymph node staging in rectal cancer patients.