Artificial Intelligence with Transformer-Based Super Resolution for Low-Light Medical Imaging

Low-light circumstances can greatly impair image quality in Medical imaging, making nuanced analyses, interpretations of features, and clinical determinations difficult. Indistinct noise and lack of lighting, as well as loss of anatomical structures, can cloud key diagnostic indicators such as lesion boundary delineation and tissue textures. This study proposes MultiHATAtt, a transformer-based super-resolution model designed to produce enhanced and cooled low-light images in the medical modality with analytics trustworthiness, interpretability, and computational efficiency. The images serve as input, where a hybrid combination of NIH Chest X-ray and BraTS dataset allows for variability in low-lighting conditions, medical modalities, and anatomical variations. Preprocessing using a Variational Denoising Network (VDN) takes place first for adaptive noise reduction without hindering delicate grained diagnostic nuance. MultiHATAtt employsHAT to enhance multi-scale contextual information, which will demonstrate its usefulness in capturing long-range dependencies through the use of MHSA and CA mechanisms highlighting clinically relevant areas. Finally, an Attention Weighted Residual Activation (AWRA) module enhances key features and ensures negligible loss of additional information and context. Super-resolution reconstruction produces high-resolution, noise-free images with interpretable attention visualizations to emphasize vital areas. Optimization is done through the AdamW optimizer with a twin loss approach that includes L1 and perceptual loss, balancing pixel-level accuracy with perceptual quality. Experimental outcomes indicate that MultiHATAtt outperforms the most advanced low-light enhancement and super-resolution techniques under PSNR, SSIM, precision, and accuracy.