Re: Most et al: Can Multimodal Large Language Models Diagnose Diabetic Retinopathy from Fundus Photos? A Quantitative Evaluation

TO THE EDITOR:

We read with great interest the recent study by Most et al¹, which evaluated the diagnostic performance of multimodal large language models (MLLMs) for detecting and grading diabetic retinopathy (DR). The authors provide a timely and rigorous quantitative assessment and, importantly, highlight the markedly reduced sensitivity for proliferative DR. As the clinical integration of MLLMs accelerates, such critical evaluations are essential for defining both the current boundaries and the future potential of these systems.

Although the quantitative findings presented by Most et al clearly delineate performance limitations, their interpretation benefits from a deeper examination of the technical foundations of MLLM image processing. The reduced sensitivity for proliferative DR may not solely reflect deficits in the models’ reasoning capabilities. Modern MLLMs rely on visual encoders that compress images into latent-space representations, a computational necessity that inevitably modifies pixel-level information. In DR screening, however, early and clinically decisive biomarkers—such as minute microaneurysms, subtle capillary nonperfusion, and delicate neovascular fronds—depend heavily on intact spatial detail and preserved contrast. Even minimal losses introduced before the reasoning stage may meaningfully compromise the detectability of these fine-grained features.

These potential losses begin in the preprocessing pipeline itself. Initial compression during storage or transmission, subsequent format conversion to portable network graphics as reported by the authors, and the internal normalization and tokenization steps executed by the MLLM all have the potential to degrade image fidelity. Evidence from computer vision underscores the clinical relevance of these distortions. Fang et al² demonstrated that compression artifacts can obscure low-contrast characteristics, reducing human ability to distinguish between synthetic and natural images. In medical imaging, Jo et al³ showed that while moderate JPEG2000 compression (up to 5:1) preserved mammographic classification performance in a deep learning model, higher compression ratios (10:1–11:1) severely impaired accuracy. These findings collectively highlight the susceptibility of task-critical diagnostic information to preprocessing-related degradation. Thus, the reduced sensitivity reported by Most et al may partially arise from upstream information loss rather than exclusively from limitations intrinsic to the MLLM architecture.

Beyond these general preprocessing considerations, an additional modality-specific factor warrants attention: the spectral characteristics of the input imagery. The Optos P200DTx system can generate both Red-Green and Red-Green-Blue composites, the latter incorporating blue light reflectance. Blue light reflectance provides enhanced visualization of inner retinal structures and is particularly effective for identifying microstructural alterations and subtle areas of nonperfusion—features highly relevant to DR.^4,5 If the data set used for MLLM analysis primarily contained red–green composites, or if the models’ internal channel weighting reduced the contribution of the blue channel, key DR biomarkers may have been less visible. The manuscript does not specify the composite type or describe how color channels were processed by the models, information that is essential for interpreting performance, particularly in mild non-proliferative diabetic retinopathy and early proliferative diabetic retinopathy where low-contrast features are most easily obscured by suboptimal spectral representation.

In summary, the work by Most et al represents an important and thoughtfully conducted contribution to the evolving field of Artificial Intelligence-assisted retinal diagnostics. Building upon their findings, we suggest future investigations explicitly examine the influence of preprocessing artifacts, color-channel handling, and wavelength composition—particularly the role of blue light reflectance—on MLLM performance. A clearer understanding of whether current limitations arise from architectural constraints, remediable input-level information loss, or a combination of both will help guide the responsible and effective advancement of these systems toward meaningful clinical integration. We commend the authors for their valuable contribution and hope these considerations further enrich the ongoing discussion.