Pradeep Kumar B P, Rangaiah Pramod K B, Augustine Robin
Department of Computer Science and Design, Atria Institute of Technology, Bengaluru 560024, India.
Microwaves in Medical Engineering Group, Division of Solid State Electronics, Department of Electrical Engineering, Uppsala University, Box 65, SE-751 03, Uppsala, Sweden.
Photodiagnosis Photodyn Ther. 2025 Aug;54:104621. doi: 10.1016/j.pdpdt.2025.104621. Epub 2025 Jun 6.
This study compares multiple image processing and deep learning methods to demonstrate an enhanced approach to glaucoma diagnosis. The approach focuses on noise reduction using median filtering and optic disc segmentation utilizing the U-Net and U-Net+ architectures. Capsule Networks were utilized for feature extraction and Extreme Learning Machines (ELM) for diagnostic classification. Three datasets were evaluated, including DRISHTI-GS, DRIONS-DB, and HRF, utilizing important parameters such as accuracy, sensitivity, and specificity. The findings revealed that median filtering reduced noise by 97.88%, with a peak signal-to-noise ratio of 44.99. U-Net beat U-Net+ in optic disc in the process of segmentation with a Dice coefficient of 0.8557, a Jaccard index of 0.7307, and higher segmentation accuracy. The suggested model has great diagnostic accuracy, scoring 99% for DRISHTI-GS, 99.5% for DRIONS-DB, and 98.5% for HRF. These findings show that using deep learning approaches can increase glaucoma diagnosis accuracy and reliability, with important implications for healthcare applications and patient outcomes.
本研究比较了多种图像处理和深度学习方法，以展示一种改进的青光眼诊断方法。该方法侧重于使用中值滤波进行降噪，并利用U-Net和U-Net+架构进行视盘分割。利用胶囊网络进行特征提取，并使用极限学习机（ELM）进行诊断分类。使用准确性、敏感性和特异性等重要参数对三个数据集（包括DRISHTI-GS、DRIONS-DB和HRF）进行了评估。研究结果显示，中值滤波将噪声降低了97.88%，峰值信噪比为44.99。在视盘分割过程中，U-Net的表现优于U-Net+，其Dice系数为0.8557，Jaccard指数为0.7307，分割精度更高。所建议的模型具有很高的诊断准确性，在DRISHTI-GS数据集上的得分为99%，在DRIONS-DB数据集上为99.5%，在HRF数据集上为98.5%。这些发现表明，使用深度学习方法可以提高青光眼诊断的准确性和可靠性，对医疗保健应用和患者预后具有重要意义。
