A multi-stage progressive multi-attention transformer for optical magnification medical images super-resolution

Optical magnification medical images are essential for doctors diagnosing diseases by analyzing patterns, textures, and regional structures. However, optical magnification medical images are often affected by various factors such as device limitations, changes in lighting, and the condition of the patient’s lesion area, resulting in low resolution and partial blurring of the obtained image results, which in turn affects subsequent diagnosis and treatment. Although super-resolution is an effective way to improve the clarity of medical images, most super-resolution models still face a series of challenges when applied to optical magnification medical images: on the one hand, most models are single-stage architecture design that cannot fully utilize multi-scale information, and on the other hand, whether the state-of-the-art methods can maintain superior performance and applicability on different medical image datasets. To address these challenges, we propose a Multi-Stage Progressive Multi-Attention Transformer (MPMAT) framework, which captures pixel information of small-, medium-, and large-scale images through a three-stage design and utilizes feature information of different scales to generate higher-quality medical images. Then, we propose a Multi-Attention Group (MAG) method, combining various attention mechanisms with a Multi-Convolution Feature Fusion Block (MCFFB). This approach can enhances the integration of local and global features and addresses the spatial and channel feature extraction strategy problem. Finally, the experimental results demonstrate that MPMAT outperforms the state-of-the-art methods by a significant margin, achieving an improvement of 0.3 dB to 1.2 dB on multiple optical magnification image datasets.