Scholarlink

<p>Glaucoma, characterized by progressive optic nerve head (ONH) damage and often linked to elevated intraocular pressure, remains a major global health concern. In imaging driven medical domain, glaucoma biomarkers are quantitative descriptors or features extracted from fundus images that act as interpretable surrogates of pathology. Automated glaucoma stage classification from fundus images is a challenging multiclass pattern-recognition task because inter-stage differences are subtle, intra-class variability is high, and labeled samples are limited. To address these pattern-analysis challenges we propose G-StageNet, a compact hybrid framework that fuses interpretable radiomic biomarkers with discriminative CNN embeddings at feature level via a lightweight, trainable gating mechanism. Fusion is supported by a projection-level triplet-loss objective that structures the transformed embedding space without updating the convolutional backbone thereby inducing moderate intra-class compactness and inter-class separation within the projected feature space. Final decisions are produced by a heterogeneous ensemble classifier evaluated under 5-fold stratified cross-validation. From a pattern analysis viewpoint, G-StageNet offers an adaptive modality-weighting strategy that balances interpretability and representational expressiveness while preserving extreme compactness for resource-constrained deployment. The proposed model attains an accuracy and AUC of 98.77% and 0.9995 respectively while remaining highly compact with a reduced model size of 0.54 MB compared to recent baseline models.</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

G-StageNet: adaptive feature fusion based hybrid learning framework for multi-class glaucoma classification from fundus imaging

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Nibedita Kalita,
  • Samir Kumar Borgohain

摘要

Glaucoma, characterized by progressive optic nerve head (ONH) damage and often linked to elevated intraocular pressure, remains a major global health concern. In imaging driven medical domain, glaucoma biomarkers are quantitative descriptors or features extracted from fundus images that act as interpretable surrogates of pathology. Automated glaucoma stage classification from fundus images is a challenging multiclass pattern-recognition task because inter-stage differences are subtle, intra-class variability is high, and labeled samples are limited. To address these pattern-analysis challenges we propose G-StageNet, a compact hybrid framework that fuses interpretable radiomic biomarkers with discriminative CNN embeddings at feature level via a lightweight, trainable gating mechanism. Fusion is supported by a projection-level triplet-loss objective that structures the transformed embedding space without updating the convolutional backbone thereby inducing moderate intra-class compactness and inter-class separation within the projected feature space. Final decisions are produced by a heterogeneous ensemble classifier evaluated under 5-fold stratified cross-validation. From a pattern analysis viewpoint, G-StageNet offers an adaptive modality-weighting strategy that balances interpretability and representational expressiveness while preserving extreme compactness for resource-constrained deployment. The proposed model attains an accuracy and AUC of 98.77% and 0.9995 respectively while remaining highly compact with a reduced model size of 0.54 MB compared to recent baseline models.