HemaViT: transformer-based deep learning for automated non-invasive anemia detection using conjunctival imaging

Abstract Anemia is a common hematological disorder that requires timely diagnosis to reduce the risk of severe health complications, particularly in resource-limited healthcare settings. This study proposes HemaViT, a transformer-based deep learning framework for automated, non-invasive detection of anemia from conjunctival images. The proposed framework integrates Dual-Attention PSPNet for accurate conjunctiva segmentation, a Feature Pyramid Network (FPN) with Multiscale Feature Exposure (MFE) for hierarchical feature extraction, a Vision Transformer (ViT) for global contextual representation, and the Improved Waterwheel Plant Algorithm (IWPA) for automated hyperparameter optimization. Experiments were conducted on the publicly available Eyes Defy Anemia dataset, which contains 1,320 conjunctival images, using stratified five-fold cross-validation. HemaViT achieved an average accuracy of 95.8%, precision of 97.6%, recall of 96.9%, F1-score of 97.2%, specificity of 98.7%, and an AUC-ROC of 0.98. Comparative experiments demonstrated that the proposed framework consistently outperformed widely used deep learning models, including ResNet50, DenseNet121, EfficientNet-B0, MobileNetV3, and a CNN + RNN hybrid model, while maintaining a favorable balance between classification performance and computational complexity. Ablation studies further confirmed the contributions of conjunctiva segmentation, multi-scale feature extraction, transformer-based contextual learning, and IWPA-driven hyperparameter optimization to the overall performance. Although additional validation on larger, more diverse clinical datasets is required, the proposed framework demonstrates strong potential for automated, non-invasive anemia screening in mobile health and resource-constrained clinical environments.