Gastric Endoscopy Image Classification Based on MobileSiT-SimAM

Gastric endoscopy image classification is an important task for the diagnosis and treatment of gastrointestinal diseases. However, existing methods based on convolutional neural networks (CNNs) or vision transformers (ViTs) have some limitations, such as insufficient feature extraction, fixed receptive field, or high computational cost. In this paper, we propose a novel method based on MobileSiT-SimAM, which integrates a lightweight ViT model (MobileSiT) and a self-adaptive attention mechanism (SimAM). MobileSiT adopts a hierarchical structure and a shifted window partitioning scheme to achieve information sharing across layers and dynamic receptive fields. SimAM can filter and weight the feature maps, enhance task-relevant features and suppress task-irrelevant features, and improve the performance of the CNN model without increasing the number of parameters. We conduct experiments on the processed Kvasir dataset, the results show that our method achieves 98.57% accuracy, which outperforms several state-of-the-art models, such as VGGNet, MobileNetV3, GoogleNet, MobileViT, and ResNet. Our method also demonstrates robustness to different image sizes and noise levels. Therefore, our method can provide an effective and efficient solution for gastric endoscopy image classification.