PViT: pooling vision transformer for active trachoma image classification

Trachoma, a preventable eye infection, poses a significant risk of blindness without timely detection and treatment. In response to this urgent need, this paper develops PViT, a novel hybrid convolutional neural network, and transformer framework that integrates an enhanced multilayer perceptron (MLP) module for trachoma image classification. To address the limitations of existing approaches in capturing critical global features, the proposed enhanced MLP module enables global information pooling and local information extraction toward compressive representation learning. Evaluations on two publicly available datasets, the active trachoma inverted eyelid image dataset, and the glaucoma OCT scan dataset, verify the effectiveness of the proposed PViT model, reaching 90% accuracy that surpasses current state-of-the-art deep learning models. Ablation studies further show the significance of the pooling technique in the enhanced MLP module.