Deep Learning for Imbalanced Gastrointestinal Image Classification: A Comparative Study of Architectural Choices

Gastrointestinal (GI) diseases refer to a wide range of disorders that affect different parts of the digestive system, including the esophagus, stomach, small intestine and large intestine. Pathological features play a crucial role in classifying gastrointestinal diseases. Early and precise diagnosis of GI tract abnormalities is essential for minimizing cancer-related mortality. While deep learning offers promising computer-aided diagnosis (CAD) capabilities, computational needs face hurdles for broader use. This study employs the Hyperkvasir dataset, which features 10,662 endoscopic images of 23 diverse classes. This study examines how architectural decisions affect the performance of deep learning models when applied to imbalanced and multi-class gastrointestinal image classification datasets. We assessed the performance of multiple CNN and transformer architectures (EfficientNet, DenseNet121, DenseNet161 MobileViT-V2, ResNet34, ResNet152, DeiT3 and GoogLeNet) concerning feature representation and gradient flow under conditions of limited data for minority classes. We further studied the influence of activation functions and loss functions. Our investigation demonstrates that DenseNet161 has the maximum performance with an accuracy of 83.03%, a macro-F1 score of 0.52 and an Area under the receiver operating characteristic Curve of 0.98, suggesting its suitability for GI image classification in resource-constrained scenarios.