Jeroen Aah Pas, Ludo C van der Zanden, Yoeri van Leeuwen, Yara Te Lechanteur, Carel B Hoyng, Thomas Theelen
We developed a preliminary nonvalidated script for cone detection in retinal AO images with support from GPT-4. Before such a script can be used in clinical research, it should undergo more fine-tuning and extensive testing. Future work should enhance the image analysis capabilities of the script and validate its results to assess the potential of AO-based cone counts as biomarkers in clinical trials.
BACKGROUND: The availability of dedicated image analysis scripts for adaptive optics (AO)-flood illumination ophthalmoscopy (FIO) is limited, especially for large-scale measurements and analyses of nonhealthy images. This limitation highlights the need for alternative approaches to facilitate automated and scalable analysis. Large language models may help develop such scripts.
OBJECTIVE: This study aimed to generate an analysis script for AO-FIO images in the R programming language using a widely available generative AI (GenAI; specifically, GPT-4) as a proof of principle for generating a functional but nonvalidated script.
METHODS: GPT-4 was used to generate an R script for the analysis of AO-FIO images. The code generated by GPT-4 was fine-tuned iteratively based on trial and error, testing the script for image preprocessing and analysis using images from 4 participants, including 1 healthy individual and 3 patients with Stargardt disease. The script code was subsequently checked for errors by another researcher who was naive to previous coding, using a different test set of AO-FIO images from 4 other participants (n=1 healthy individual and n=3 patients with Stargardt disease). The cone counts from 5 AO image snippets were compared with the counts independently recorded by 2 human graders and those generated by pre-existing AO analysis software that was trained on healthy participants.
RESULTS: After 54 iterations of instructions, a functional R script for the analysis of AO-FIO images was generated. The script identified and quantified blobs.
CONCLUSIONS: We developed a preliminary nonvalidated script for cone detection in retinal AO images with support from GPT-4. Before such a script can be used in clinical research, it should undergo more fine-tuning and extensive testing. Future work should enhance the image analysis capabilities of the script and validate its results to assess the potential of AO-based cone counts as biomarkers in clinical trials.