FGMG-CAViT: An adaptive fuzzy contextual multi-granular vision-based learning model for X-ray imagery enhancement

Radiography, also known as X-ray imaging, is widely used to identify fractures, diagnose pulmonary conditions such as pneumonia, evaluate dental pathologies, and conduct relevant examinations. It helps produce detailed images that can aid in the assessment and management of various medical conditions. However, numerous challenges, such as noise and artefacts, can obscure very important diagnostic information. Diagnostic accuracy can reduce due to technical issues during image capture and processing, as well as misalignment, leading to added noise and distortion in images. Noise can stem from many sources such as: secondary radiation, equipment issues, digitization errors, etc. Most of these can be reduced through the use of different kinds of filters. Such problems can result in low-quality captures. Contextual enhancement techniques can significantly improve clarity and contrast. This is especially beneficial in podiatric imaging and for X-ray imaging of patients with respiratory issues. X-ray images can provide valuable information. These often have noise and single granularity levels for feature extraction. Multiscale analysis techniques are known to enhance feature extraction by capturing varying granularities, thereby helping address intra-class variance. Redundancy may be introduced due to multi-granular features, but it can help focus on relevant correlations that enhance contextual content. Fuzzy logic can further ease the aggregation of essential features based on their membership functions. We propose "FGMG-CAViT," a fuzzy-guided, multi-granularity, contextually aware, vision-based deep-learning model for enhancing X-ray images. The main contribution of the presented model includes (i) A multi-granularity feature extraction method that allows examination of semantic scales, (ii) the use of universal fuzzy features to combine significant characteristics across varying granularities through fuzzy logic, and (iii) Fuzzy-guided techniques that assign varying relevance to retrieved features, improving overall feature descriptions for deep learning models. In this paper, we present the findings from the implementation of the proposed FGMG-CAViT methodology, for image enhancement across multiple datasets. Comparison is performed against twelve other image processing methods. For evaluation purposes, traditional techniques, such as histogram equalization and its variations, as well as advanced approaches, including deep neural networks and transform domain methods were also tested with alongside. FGMG-CAViT was also benchmarked against prominent state-of-the-art methods like Brightness Preserving Dynamic HE, CLAHE, Multi-level CLAHE, Wavelet transformation, Contourlet, CNN, U-Net, and ResNet. Various X-ray datasets, including Chest X-rays, Podiatry, Dental, Orthopantomography, and Spinal images, were used for testing. FGMG-CAViT consistently outperformed the alternatives in metrics such as PSNR, CII, WC, Normalized MC, EME, EMEE, UIQI, and VIF, indicating superior quality and contrast enhancement. Notably, in the Podiatry dataset, FGMG-CAViT achieved a PSNR of 38.902 and a CII of 4.752, outperforming other traditional methods. The technique also demonstrated the highest PSNR and CII for Orthopantomography images, reinforcing its efficacy in enhancing image quality. While traditional methods provide satisfactory results, they often fall short of advanced techniques, such as FGMG-CAViT, in meeting specific diagnostic needs. This study highlights the possible transformative potential of FGMG-CAViT in medical imaging, consistently demonstrating enhanced performance across varied applications and datasets, ultimately improving diagnostic accuracy. The proposed methodology presents findings from experiments utilizing the FGMG-CAViT solution for image enhancement across multiple datasets, comparing it against twelve other image processing methods. The evaluation encompasses traditional techniques and advanced approaches, including deep neural networks and transform domain methods. FGMG-CAViT was benchmarked against prominent state-of-the-art methods, including brightness and contrast enhancement methods, transform-domain methods, certain learning-based methods, and fuzzy logic-based methods.