Multi-task deep learning framework for enhancing Mayo endoscopic score classification in ulcerative colitis

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Ulcerative colitis (UC) endoscopic image classification is challenging due to imbalanced datasets.
• Accurate positive predictions are crucial, especially with increasing disease severity.

Purpose of the Study:

• To propose a multi-task learning (MTL) framework for UC endoscopic image classification.
• To address data imbalance issues and improve detection of advanced disease stages.

Main Methods:

• Developed an MTL framework inspired by human brain processing.
• Evaluated the framework on UC endoscopic images, focusing on disease stage classification.
• Utilized DenseNet121 and MobileNet-v3-large as model backbones.

Main Results:

• The MTL framework effectively mitigated data imbalance issues.
• Classification performance was enhanced for severe UC disease stages (Mayo scores 2 and 3).
• DenseNet121 showed superior performance; MobileNet-v3-large also demonstrated gains, indicating framework versatility.

Conclusions:

• MTL-based computer-aided diagnosis aids in accurate identification of critical UC stages.
• This supports timely treatment decisions and reduces underdiagnosis.
• Future work should explore integrating multiple CNN models to further improve accuracy.