Scholarlink

<p>Eye disease is one of the most conspicuous reasons of optical diminishing and, in certain cases, the prime cause of complete sightlessness. So, there is an urgent need to innovate systems that are capable of spotting eye sicknesses such as glaucoma, cataracts and diabetic retinopathy (DR). To solve this, motivating the necessity for accurate systematized assessment systems, we have advanced a two-backbone deep learning (DL) structure that systematically conglomerates EfficientNetB0 with three harmonizing architectures: ResNet50, InceptionV3, and AlexNet, exercising four different fusion strategies: concatenation, element-wise summation, weighted and majority voting. Using a stable dataset of 4,217 High Resolution retina images across the four symptomatic classes, we have educated and evaluated 12 fusion model arrangements to recognize ideal feature assimilation methodologies. Internal validation exhibited that concatenation fusion attained the great baseline performance, with EfficientNetB0 + ResNet50 (Exp01) triumph 95.26% accuracy (AUROC: 0.993) and EfficientNetB0 + InceptionV3 getting 94.79% accuracy (AUROC: 0.989). External validation on 400 images from Messidor-2 and ODIR datasets disclosed significant performance increases, with accuracies oscillating from 94.99% to 97.99% and AUROC values between 0.991 and 0.999, confirmatory authentic cross-dataset generalization rather than overfitting. Weighted and sum fusion strategies evidenced particularly effective for the external dataset, with EfficientNetB0 + ResNet50 (Exp03) weighted fusion reaching 97.49% accuracy, and EfficientNetB0 + AlexNet (Exp09) sum fusion accomplished an MCC of 0.980, signifying that intelligent feature combination can counterbalance for architectural boundaries and enhance domain robustness. Class-wise heatmap analysis displayed that while DR triumphed near-perfect detection across all configurations, glaucoma detection improved considerably with weighted, sum, and voting fusion, and reduced misclassifications with normal eye images. Explainability and interpretability examination using Score-CAM (disclosed steady anatomical attention patterns across both datasets: glaucoma models concentrated on the optic nerve head and peripapillary area, DR models emphasized macular zones and vascular structures, and cataract models focused on lens denseness all aligning with traditional medical checkups standards and proven that model decisions are based in pathophysiological relevant features rather than spurious connections. Our findings are that the dual-backbone fusion architectures with improved feature integration strategies, remarkably weighted and sum fusion, produce not only higher analytical accuracy but also strong simplification and clear quick decision-making essential for actual clinical implementation across various imaging locations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-class eye disease classification using deep learning EfficientNetB0 fusion techniques

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月13日

查看全文
  • Uday Kumar Sah,
  • Jyotir Moy Chatterjee,
  • R. Sujatha

摘要

Eye disease is one of the most conspicuous reasons of optical diminishing and, in certain cases, the prime cause of complete sightlessness. So, there is an urgent need to innovate systems that are capable of spotting eye sicknesses such as glaucoma, cataracts and diabetic retinopathy (DR). To solve this, motivating the necessity for accurate systematized assessment systems, we have advanced a two-backbone deep learning (DL) structure that systematically conglomerates EfficientNetB0 with three harmonizing architectures: ResNet50, InceptionV3, and AlexNet, exercising four different fusion strategies: concatenation, element-wise summation, weighted and majority voting. Using a stable dataset of 4,217 High Resolution retina images across the four symptomatic classes, we have educated and evaluated 12 fusion model arrangements to recognize ideal feature assimilation methodologies. Internal validation exhibited that concatenation fusion attained the great baseline performance, with EfficientNetB0 + ResNet50 (Exp01) triumph 95.26% accuracy (AUROC: 0.993) and EfficientNetB0 + InceptionV3 getting 94.79% accuracy (AUROC: 0.989). External validation on 400 images from Messidor-2 and ODIR datasets disclosed significant performance increases, with accuracies oscillating from 94.99% to 97.99% and AUROC values between 0.991 and 0.999, confirmatory authentic cross-dataset generalization rather than overfitting. Weighted and sum fusion strategies evidenced particularly effective for the external dataset, with EfficientNetB0 + ResNet50 (Exp03) weighted fusion reaching 97.49% accuracy, and EfficientNetB0 + AlexNet (Exp09) sum fusion accomplished an MCC of 0.980, signifying that intelligent feature combination can counterbalance for architectural boundaries and enhance domain robustness. Class-wise heatmap analysis displayed that while DR triumphed near-perfect detection across all configurations, glaucoma detection improved considerably with weighted, sum, and voting fusion, and reduced misclassifications with normal eye images. Explainability and interpretability examination using Score-CAM (disclosed steady anatomical attention patterns across both datasets: glaucoma models concentrated on the optic nerve head and peripapillary area, DR models emphasized macular zones and vascular structures, and cataract models focused on lens denseness all aligning with traditional medical checkups standards and proven that model decisions are based in pathophysiological relevant features rather than spurious connections. Our findings are that the dual-backbone fusion architectures with improved feature integration strategies, remarkably weighted and sum fusion, produce not only higher analytical accuracy but also strong simplification and clear quick decision-making essential for actual clinical implementation across various imaging locations.