Gastrointestinal Lesion Detection Using Ensemble Deep Learning Through Global Contextual Information.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Subtle mucosal abnormalities in small bowel Crohn's disease (SBCD) and other gastrointestinal lesions are challenging to detect via manual endoscopic interpretation.
• Current methods are labor-intensive, time-consuming, and subject to clinician variability.
• Existing deep learning models may miss critical details due to shallow feature extraction.

Purpose of the Study:

• To develop a generalizable ensemble deep learning framework for accurate gastrointestinal lesion detection.
• To improve the identification of subtle pathological patterns resembling SBCD-associated abnormalities.
• To overcome the limitations of classical Convolutional Neural Networks (CNNs) in capturing complex lesion features.

Main Methods:

• An ensemble deep learning framework combining EfficientNetB5, MobileNetV2, and multi-head self-attention (MHSA) was developed.
• EfficientNetB5 was used for detailed hierarchical feature extraction.
• MobileNetV2 enhanced spatial representation, and MHSA improved global feature correlation.

Main Results:

• The model achieved high classification accuracies of 99.25% on the Kvasir dataset and 98.86% on the Kaither dataset.
• Performance was evaluated on two publicly available DBE datasets.
• Results were compared favorably against four state-of-the-art methods.

Conclusions:

• The proposed ensemble deep learning framework effectively detects subtle gastrointestinal lesions.
• This approach offers a more accurate and potentially less subjective alternative to manual endoscopic analysis.
• The model demonstrates significant potential for improving the diagnosis of conditions like SBCD.