Explainable Artificial Intelligence for the Mayo Endoscopic Score in Ulcerative Colitis.

Area of Science:

• Gastroenterology
• Artificial Intelligence
• Medical Imaging

Background:

• The Mayo endoscopic score (MES) is crucial for assessing ulcerative colitis (UC) severity and guiding treatment.
• Current deep learning (DL) models for MES lack transparency, hindering clinical trust and adoption.
• Explainable AI (XAI) offers a solution to enhance the interpretability of DL models in medical diagnostics.

Purpose of the Study:

• To develop and evaluate explainable AI (XAI) models for determining the MES in UC patients.
• To compare the performance of XAI models against traditional DL models and human endoscopists.
• To investigate the impact of AI assistance on endoscopist performance in MES interpretation.

Main Methods:

• A retrospective multicenter study utilized 2,600 white-light endoscopic images for training DL models.
• Two approaches were employed: traditional black-box models and explainable AI (XAI) models.
• Models were validated on three external datasets and compared with endoscopist assessments, using Grad-CAM for visual explainability.

Main Results:

• A Xception-backboned XAI model achieved high accuracy (0.910) and correlation coefficients on external datasets, outperforming other models and endoscopists.
• AI assistance significantly improved endoscopist performance, increasing senior endoscopist accuracy from 0.890 to 0.930 and junior endoscopist accuracy from 0.810 to 0.880.
• Similar performance improvements were observed across all three external test datasets, confirming model robustness.

Conclusions:

• Employing an explainable AI framework enhances DL model performance and transparency in MES assessment for UC.
• XAI tools can augment endoscopist capabilities, leading to more accurate and reliable interpretations of endoscopic findings in UC.
• The integration of XAI holds promise for improving clinical decision-making and patient outcomes in ulcerative colitis management.