Decomposing Cross-Modality Failure in Diabetic Retinopathy Grading

Deep learning detects referable diabetic retinopathy (DR) from colour fundus photography at a level comparable with specialists, but such models are rarely tested across a change of imaging modality. We evaluate a colour-trained DR grader zero-shot on near-infrared reflectance images acquired by confocal scanning laser ophthalmoscopy, and report two findings. First, the transfer fails without any detectable signal in the model output: 96.1% of images collapse to the no-DR class while mean confidence rises from 0.619 to 0.675 (Cohen’s d = 0.41), so an abstention policy calibrated in domain would pass the collapsed predictions through. The apparent severe-grade tail correlates with image brightness and is unrelated to pathology. Second, and centrally, the failure is not monolithic. A deterministic green-channel intensity transform, which by construction cannot synthesise structure, recovers substantial transferable signal: collapse falls to 53.2%, grade-to-thickness concordance flips from -0.492 to +0.332, and sensitivity for a macular fluid composite rises from 5.2% to 65.2%, at a cost of 0.020 in in-domain validation quadratic weighted kappa. Because the transform cannot fabricate structure, the recovered signal is attributable to distributional alignment alone, which places a measurable lower bound on the preprocessing-addressable share of the gap and bounds the remainder. The dominant barrier is representational: single-wavelength reflectance replicated across three channels and normalised with colour-image statistics occupies a region of input space the encoder never encountered in training, and the chromatic contribution is comparatively small. We validate indirectly against clinical proxies, since no public near-infrared dataset carries ground-truth DR grades, and we are explicit that this constitutes convergent evidence and does not amount to proof.