LightGastroFormer: a lightweight multi-resolution transformer for gastrointestinal disease classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Automated analysis of gastrointestinal (GI) images is crucial for early disease diagnosis and reducing physician workload during endoscopic procedures.
• Current deep learning models struggle with large, imbalanced datasets, failing to capture both long-range context and fine-grained local patterns effectively.

Purpose of the Study:

• To introduce LightGastroFormer, a novel, lightweight transformer-based architecture for accurate and efficient GI disease categorization.
• To address the limitations of existing deep learning methods in handling complex GI image datasets.

Main Methods:

• Developed LightGastroFormer, a transformer architecture featuring a gated feed-forward network, efficient self-attention, and a multi-resolution patchwise tokenizer.
• Evaluated the model on three public benchmarks: Kvasir v1, Kvasir v2, and the Kvasir-Capsule dataset.

Main Results:

• LightGastroFormer achieved high performance across all datasets, matching or exceeding state-of-the-art methods.
• Achieved 0.94 accuracy on Kvasir v1, 0.95 on Kvasir v2, and 0.97 accuracy with a 0.97 F1-score on the Kvasir-Capsule dataset, without explicit data balancing.
• Demonstrated effectiveness of architectural components and robustness under class imbalance.

Conclusions:

• LightGastroFormer offers a robust, efficient, and accurate solution for GI disease classification, suitable for clinical deployment.
• The model's lightweight design (6.42 million parameters) balances performance with computational efficiency for real-world applications.