Answers are grounded in this paper and cite their sources.
Critical narrative review explores lesion-location-change reasoning in ophthalmic vision-language models, highlighting its implications for clinical care.
Artificial intelligence (AI), machine learning (ML), and deep learning (DL) have already produced clinically deployed ophthalmic systems for diabetic retinopathy (DR) screening and high-performing algorithms for image-level classification and lesion segmentation [1-8]. However, clinical retinal interpretation is not equivalent to assigning a global disease label. A clinically useful reader must associate lesions with anatomical locations, compare the same location across time, and express the finding in a way that can guide treatment or referral. To review evidence that the major unsolved problem for present ophthalmic AI is not whether an image can be globally classified as DR, but whether vision-language models (VLMs), large vision-language models (LVLMs), and multimodal large language models (MLLMs) can perform grounded lesion-location-change reasoning at the level required for clinical care [9-25]. We performed a critical narrative review of peer-reviewed ophthalmology, medical imaging, and AI literature, supplemented by regulatory documents. The review focused on DR, diabetic macular edema (DME), hard exudates, fundus foundation models, medical VLMs, visual grounding, hallucination, and reinforcement learning [6-7, 14-39]. Non-peer-reviewed news and trade-magazine sources were removed from the evidentiary base. Image-level and screening systems have reached high performance in selected settings, as shown by JAMA and FDA-cleared autonomous DR work [4-7]. Pixel-level lesion detection can also be achieved when lesion masks and anatomical labels are available, as shown by IDRiD and DeepDR [8, 40-41]. In contrast, current VLMs perform inconsistently on ophthalmic multimodal image analysis and remain vulnerable to spatial-reasoning errors, visual hallucination, and weak anatomical localization [14-25]. The clinical example of'increased hard exudates in the macula' requires foveal/macular localization, hard-exudate detection, registration across visits, quantification, and text generation as a single grounded chain [40-45]. A reinforcement learning (RL) approach cannot solve this by rewarding final disease labels; it requires verifiable dense rewards, including fovea-localization error, lesion Dice/IoU, lesion-to-fovea distance, and longitudinal change metrics [24-25, 35-36]. The next clinically decisive frontier for ophthalmic AI is grounded clinical reading: the ability of a VLM to bind lesion identity, anatomical location, and longitudinal change into a verifiable statement. This review proposes lesion-location-change reasoning as the critical benchmark for future ophthalmic VLMs. Artificial intelligence systems can already screen retinal photographs for diabetic retinopathy in defined workflows. This does not mean that they read retinal images like ophthalmologists. A clinically useful reader must identify the lesion, identify where it is located, compare the same location over time, and state whether it has worsened. The sentence 'hard exudates have increased in the macula' is therefore a stringent test of ophthalmic vision-language models. This review proposes that future ophthalmic AI systems should be evaluated not only by diagnosis-level accuracy but also by fovea localization, lesion segmentation, lesion-to-fovea relationships, longitudinal change, and text-image consistency.
No takes yet. Share an insight, caveat, or question.
Tabuchi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: