Enhancing Gastrointestinal Disease Diagnosis Using Fine-Tuned MobileNetV2

Gastrointestinal (GI) diseases lead to disturbances in the gut health, causing conditions like ulcers, polyps, and cancers that contribute to overall health status. Their early detection is essential to prevent severe outcomes, and provide effective treatment. Traditional endoscopic image interpretation is time-intensive and heavily reliant on medical expertise, leaving room for manual errors. Deep learning models such as VGG16 and ResNet34 are prone to overfitting, which leads to poor generalization and they require high computational demands, compromising their effectiveness in real-world scenarios. This study utilizes MobileNetV2, a lightweight deep learning architecture, for GI disease classification. The model was trained on two datasets- Kvasir V2 dataset containing 8,000 endoscopic images and Kvasir containing 4,000 images across eight GI disease classes. This study employs several refined techniques, including label smoothing, L2 regularization, data augmentation and cosine annealing learning rate scheduling, to enhance model adaptability and prevent overfitting. MobileNetV2 demonstrated superior performance to other models, achieving 95% accuracy on Kvasir and 94% on Kvasir V2. Findings emphasize MobileNetV2's potential for real-world integration and strong generalization capability, offering a scalable and dependable AI-driven solution to assist healthcare professionals in accurate and timely GI disease diagnosis.