UCFNNet: Ulcerative colitis evaluation based on fine-grained lesion learner and noise suppression gating

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Ulcerative colitis (UC) is a chronic inflammatory bowel disease with unknown etiology, leading to significant patient morbidity.
• Current UC treatment selection relies on disease severity and location, necessitating accurate evaluation methods.
• Automated analysis of endoscopic images is crucial for guiding UC treatment and enabling early prevention strategies.

Purpose of the Study:

• To develop a fully automated deep learning network for evaluating ulcerative colitis (UC) severity from endoscopic images.
• To enhance the accuracy of lesion detection and feature extraction in UC endoscopic images.
• To provide a robust tool for clinical decision-making in UC management.

Main Methods:

• Proposed UCFNNet (ulcerative colitis evaluation based on fine-grained lesion learner and noise suppression gating).
• Integrated a fine-grained lesion feature learner (FG-LF Learner) with Softmax category prediction (SCP) for improved small lesion accuracy.
• Employed a graph convolutional feature combiner (GCFC) to reduce feature loss and a noise suppression gating (NS gating) technique for feature prioritization.

Main Results:

• Achieved high performance on a privately-collected dataset: Accuracy (ACC) 89.57%, Matthews correlation coefficient (MCC) 85.52%, F1-score 89.78%.
• Demonstrated strong results on a publicly-available dataset: ACC 85.47%, MCC 80.42%, F1-score 84.53%.
• Outperformed existing state-of-the-art techniques in UC evaluation.

Conclusions:

• The proposed UCFNNet model, with its novel algorithmic modules, significantly surpasses current methods in UC evaluation.
• The model exhibits superior performance compared to traditional machine learning and existing deep learning approaches.
• UCFNNet demonstrates good interpretability and holds significant potential for clinical application in UC management.