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

We thank Henriques et al for the thoughtful comments in response to our recent article¹ and the methodologies used. The following reply aims to address the 2 key questions raised.

The first question, regarding image quality, suggests that the accuracy of multimodal large language models (MLLMs) could be reduced secondary to image quality losses from reformatting. We agree that it is a reasonable assumption that higher MLLM accuracy might be achieved with the highest possible quality images. In our study, however, image reformatting was necessitated by MLLM acceptance of only certain file types, including Portable Networks Graphics, Joint Photographic Experts Group, WebP, and Graphics Interchange Format for Chat GPT-4o. Portable Networks Graphics was selected as a reasonable formatting option, given that it was accepted by all of the tested MLLMs, and compression ratios are comparable to Joint Photographic Experts Group.² Future MLLM models may accept additional file types; however, the same formatting limitations presently remain for some of the tested models.

It is also worth noting that our human graders were shown identically formatted images to establish ground truth image grades for the study. If a pathologic image feature was theoretically obscured by reformatting, the ground truth human grade should reflect the absence of this feature. Notably, graders also had near-perfect agreement for binary disease classification. Although grades for disease severity had a lower consensus (as expected for a more complex question), consensus severity grades were still reached for 94.5% of eyes. This suggests that the images could be reasonably graded at the level of quality preserved in our Portable Networks Graphics images.

The second question centers on the spectral characteristics of images captured by the Optos device. The study was conducted with the prior-generation Optos P200Tx. Our institution was one of the first to have the new 3-color Optos,³ and we are very familiar with the advantages of the 2-color red–green–blue (RGB) Optos. In general, the spectral distribution of each color channel is not equal in the imaging of the eye. In our prior work, we determined that the red and green channels contain more detail than the blue channel. This is common knowledge in the ophthalmology field and may be the reason why Optos initially released only red–green (RG) devices.

We have chosen to concentrate our model on the RG components because the majority of Optos devices in the world are still the RG model. The 3-color RGB model has only recently been released, and most clinics may not have it. We agree that this may be a limitation of the model, and the model can be improved with the inclusion of RGB Optos images, but the majority of historical retinal images in most clinics will be limited to the RG Optos, and these images are the most likely available for training the MLLMs.

We thank the authors of the letter for their insightful additions to this discussion, and hope our reply clarifies the authors’ comments.