Treffer: Glaucoma detection and severity classification based on glaucoattent net framework.

Global blindness due to glaucoma necessitates advanced diagnostics. Automated glaucoma detection models, while introduced, grapple with discerning subtle optic disc (OD) and optic cup (OC) changes. Challenges arise in intricate segmentation due to structural variations. Moreover, many existing works fixate on binary classification, lacking nuanced assessments, highlighting the need for more sophisticated approaches. To tackle these challenges, we introduce a framework called GlaucoAttent Net for glaucoma detection, aiming to effectively segment the OD and OC while assessing severity. The foundation of robust model training lies in diverse datasets—ORIGA, REFUGE and G1020. Initially, a glaucoma feature enhanced pre-processing stage preserves fundus image details through Non-Local Means (NLM) denoising, Contrast Limited Adaptive Histogram Equalization (CLAHE) and Adaptive Median Filtering. It retains fine structures, texture patterns, OD morphology, intensity gradients and localized features essential for segmentation. Within the core engine GlaucoAttent Net, which is a cascaded U-Net, the encoder path efficiently extracts feature at two levels to create hierarchical feature representations essential for segmenting the OD. The decoder employs a novel attention mechanism to create a probability mask for OD segmentation followed by OC segmentation by refining features. At the last stage of decoder, merging the probability masks for accurate delineation of the OD and OC boundaries in the image. Additionally, a classification system analyses segmentation outputs like cup-to-disc ratio, areas, diameters and other features to categorize eyes as Healthy, Highly Glaucomatous, Less Glaucomatous and Moderately Glaucomatous. Efficacy metrics showcase outstanding performance: 99.67% accuracy, 99.20% precision, 99.50% recall and a 99.35% F-score, solidifying the methodology's potential for transformative impact in advancing clinical diagnostics. [ABSTRACT FROM AUTHOR]

Copyright of International Journal of Machine Learning & Cybernetics is the property of Springer Nature and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)