Comparative Analysis of Deep Neural Networks for Automated Ulcerative Colitis Severity Assessment.

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Ulcerative colitis (UC) is a chronic inflammatory bowel disease requiring accurate severity assessment for effective treatment.
• The Mayo Endoscopic Score (MES) is a key diagnostic tool, but its subjective nature leads to inconsistencies.
• Deep learning offers potential for objective and standardized UC severity evaluation.

Purpose of the Study:

• To compare the performance of deep neural networks for automated MES classification in UC.
• To identify the most effective deep learning models for grading UC endoscopic disease severity.

Main Methods:

• Utilized publicly available endoscopic images from UC patients.
• Evaluated multiple state-of-the-art deep neural network architectures for automated MES classification.
• Calculated F1 score, accuracy, recall, and precision, with statistical analysis for significance.

Main Results:

• VGG19 demonstrated strong performance with a QWK score of 0.876 and macro-averaged F1 score of 0.7528.
• Top-performing models showed minimal performance differences, indicating deployment requirements should guide selection.
• Deep learning models achieved automated classification of UC severity.

Conclusions:

• State-of-the-art deep neural networks can effectively automate UC severity classification.
• Simpler network architectures achieved competitive results, challenging the need for larger models for better clinical outcomes.
• Automated endoscopic evaluation holds promise for more consistent UC management.