Contour-guided attention across YOLO architectures for optic structure detection

BACKGROUND: Glaucoma is a leading cause of irreversible blindness, and the Cup-to-Disc Ratio used to assess it requires localizing the optic disc and faintly bounded cup. Attention modules are frequently proposed for this task, but of six such studies meeting a pre-stated rule, none reported their number of training runs or any run-to-run variability. Methods. We evaluate a contour-guided attention module (ContourCBAM) whose spatial map comes from a fixed, non-learned contour operator applied to the input image rather than an intermediate feature map, refined by zero-initialized convolutions and injected at P3, P4 and P5 so that each block is exactly the identity at initialization. We measure the seed-noise floor over 310 baseline runs across seven YOLOv9, YOLOv12 and YOLO26 architectures, then run pre-registered, fixed-sample, fresh-seed paired comparisons and zero-shot evaluation on REFUGE. Results. The seed alone moves baseline mAP50 across a range of $0.051$, and seed matching across arms gives essentially no variance reduction, so ten paired seeds resolve only +0.006 to +0.018 mAP50, the order of gains those studies attribute to attention. Against that floor the module improved YOLOv12m in two pre-registered blocks of 30 fresh-seed pairs (ΔmAP50 =+0.0033, p=0.0079 and +0.0025, p=0.0439; ΔmAP50-95 null in both), entirely optic-cup driven and confined to IoU 0.50, where cup AP is near 0.97 against 0.03 at IoU 0.90. Conclusion. The module yields a small, replicated, optic-cup-specific gain at IoU 0.5 on one architecture, with no evidence of improved boundary precision and none of generality. No clinical benefit follows: on a box-derived cup-to-disc-ratio proxy the module is a precise null. The noise floor is the more transferable result, fixing the sample sizes at which such gains become checkable and showing that single-run and few-seed comparisons cannot support them.