HMSGA-Net: Multi-Scale Gated Attention with Dual ConvNeXt Ensembling for Glaucoma Classification from Fundus Images

Glaucoma is a major cause of irreversible blindness,
and early screening from fundus photographs is essential because
structural damage may progress before noticeable visual symp-
toms occur. This paper proposes HMSGA-Net, a Hybrid Multi-
Scale Gated Attention Network for automated glaucoma classi-
fication. The framework combines a ConvNeXt-Small primary
branch and a ConvNeXt-Tiny auxiliary branch, each equipped
with an HMSGA module comprising parallel 1 × 1, 3 × 3, and
5 × 5 depth-wise convolutions, Efficient Channel Attention, dual
sigmoid gating, and residual feature refinement. The two branch
probabilities are fused using a validation-selected ensemble
weight and decision threshold. Integrated Gradients and Occlu-
sion Sensitivity are employed to assess whether predictions focus
on clinically relevant optic-disc and neuroretinal-rim regions. On
the ACRIMA dataset, the best seed-42 configuration achieved
98.84% accuracy, 98.97% F1-score, 99.72% AUC-ROC, and
100% glaucoma recall. Across three independent splits, HMSGA-
Net obtained mean accuracy and AUC-ROC values of 94.14%
and 97.28%, respectively, demonstrating strong discrimination
while also revealing sensitivity to data partitioning.