A trustworthy evidential transformer with linear fusion and Gram-matrix distillation for efficient optic disc and cup segmentation

Accurate glaucoma screening relies on the vertical cup-to-disc ratio (VCDR), but automation remains challenging. A critical trade-off between computational efficiency, segmentation accuracy, and model trustworthiness hinders clinical deployment. Vision Transformers offer global context but suffer from quadratic complexity ( O ( N 2 ) ), while standard Convolutional Neural Networks lack explicit mechanisms for uncertainty quantification (UQ). To bridge this gap, we propose EviG-Former (Evidential Gram Transformer), a highly efficient student network trained via Teacher-Student distillation. Our framework combines Evidential Deep Learning (EDL) with Linear Multi-Scale Token Fusion (L-MSTF). We designed an Uncertainty-Gated Feature Fusion (UGF) module that uses epistemic uncertainty to guide gradient flow, focusing on ambiguous regions like faint vessel boundaries. Additionally, Gram-Matrix Consistency Distillation ensures the lightweight student maintains topological integrity. We validated EviG-Former on RIM-ONE-r3, Drishti-GS, ORIGA, and REFUGE datasets. It achieved highly competitive Optic Cup segmentation Dice scores (0.9004, 0.9423, 0.8916, and 0.9005, respectively) when evaluated against thirteen leading baselines. Crucially, this performance requires only 0.69M parameters (up to 178 × fewer than recent high-capacity networks) while sustaining real-time processing ( ≈ 60.7 FPS). Furthermore, the estimated VCDR mean absolute error falls below the 0.05 clinical threshold on three datasets. By uniting diagnostic precision, edge efficiency, and calibrated uncertainty mapping, EviG-Former effectively addresses the traditional capacity-accuracy-trustworthiness trade-off.