Guided sparse decomposition with anisotropic fusion for medical image enhancement

Medical imaging plays a crucial role in modern healthcare by providing essential insights for accurate diagnosis and effective treatment planning across a variety of diseases. However, enhancing these images while preserving critical structural details presents significant challenges, as existing techniques often compromise vital information during the enhancement process. This study aimed to develop an efficient and artifact-free image enhancement method to improve both tonal clarity and detail preservation in medical images, thereby enhancing their diagnostic utility.

We proposed a novel method, named guided sparse decomposition with anisotropic fusion (GSDAF), which improves upon the traditional guided image filter (GIF) by introducing a weighted sparse regression model to reduce luminance halos and an anisotropic coefficient fusion strategy to mitigate detail halos. We applied GSDAF to enhance the detail and tone of medical images. We evaluated the proposed method using the images from the coronavirus disease of 2019 (COVID-19) chest X-ray dataset and the Child Heart and Health Study in England-Database 1 (CHASE-DB1) dataset.

In terms of tone enhancement, GSDAF achieved the best spatial-spectral entropy-based quality (SSEQ; 26.62) and the third best convolutional neural network image quality assessment (CNNIQA; 30.70) on the CHASE-DB1 dataset. In terms of detail enhancement, GSDAF achieved the best CNNIQA (18.72) and the second best SSEQ (23.76) on the COVID-19 dataset; it achieved the best SSEQ (21.76) and the second best CNNIQA (25.55) on the normal subset; it achieved the second best SSEQ (25.97) and CNNIQA (25.29) on the viral pneumonia subset. In terms of efficiency, it processed 720P color images in 0.032 seconds on a modern graphics processing unit (GPU).

The proposed GSDAF effectively enhances medical images while preserving critical structural details. GSDAF addresses the limitations of traditional enhancement techniques. Our findings suggest that this approach can significantly improve the quality of medical imaging, thereby supporting better diagnostic and treatment outcomes in healthcare.