先进大语言模型在视网膜影像诊断准确性中的深度分析
该研究评估了八种先进多模态大语言模型对100张代表不同眼科疾病眼底图像的诊断能力,采用验证图表评估诊断准确性、特异性、敏感性、一致性、相关性和解释质量。GPT-4.5获得最高总体诊断准确性(65.0%),其次为Gemini 2.0 Pro(63.0%);孔源性病变识别最准确,近视性黄斑病变最具挑战性(平均准确性21.8%)。结论认为当前LLMs在眼科图像解读中展现出有前景但有限的能力。
要点
- GPT-4.5总体诊断准确性最高(65.0%),Gemini 2.0 Pro次之(63.0%)。
- 孔源性病变识别最准确(Gemini 2.0 Pro 81.3%,GPT-4.5 75.0%),近视性黄斑病变最具挑战性(平均21.8%)。
- 表现较差的模型常拒绝提供诊断,拒绝率从8.0%(Claude 3.7 Sonnet)到19.0%(Mixtral 8×22B)。
核心结论
- 当前LLMs在眼科图像解读中展现出有前景但有限的能力。
- 常见病变如视网膜脱离和年龄相关性黄斑变性表现中等良好,但罕见病变、近视性病变和复杂血管疾病仍存在显著挑战。
- GPT-4.5与Gemini 2.0 Pro性能竞争,各自在不同病理类别中表现优异,提示利用其互补优势可能提供更好的诊断支持。
研究设计与样本
- 评估八种领先多模态LLMs(GPT-4.5、Claude 3.7 Sonnet、Grok-2、Deepseek Cognition V2、Qwen2 72B、Gemini 2.0 Pro、Llama 3 405B和Mixtral 8×22B)对100张代表不同眼科疾病的眼底图像的解读能力。
- 使用验证图表评估诊断准确性、特异性、敏感性、一致性、相关性和解释质量。
主要发现
- GPT-4.5总体诊断准确性最高(65.0%),Gemini 2.0 Pro次之(63.0%)。
- 孔源性病变识别最准确(Gemini 2.0 Pro:81.3%,GPT-4.5:75.0%),近视性黄斑病变最具挑战性(平均准确性21.8%)。
- 其余模型表现显著较差:Deepseek Cognition V2(52.0%)、Claude 3.7 Sonnet(52.0%)、Qwen2 72B(49.0%)、Llama 3 405B(48.0%)、Grok-2(47.0%)和Mixtral 8×22B(46.0%)。
- 表现较差的模型常拒绝提供诊断,拒绝率从8.0%(Claude 3.7 Sonnet)到19.0%(Mixtral 8×22B)。
关键数值与推荐
- GPT-4.5总体诊断准确性65.0%,Gemini 2.0 Pro 63.0%。
- 孔源性病变识别准确性:Gemini 2.0 Pro 81.3%,GPT-4.5 75.0%。
- 近视性黄斑病变平均准确性21.8%。
- 其他模型诊断准确性:Deepseek Cognition V2 52.0%,Claude 3.7 Sonnet 52.0%,Qwen2 72B 49.0%,Llama 3 405B 48.0%,Grok-2 47.0%,Mixtral 8×22B 46.0%。
- 拒绝率范围:8.0%(Claude 3.7 Sonnet)至19.0%(Mixtral 8×22B)。
临床怎么用
- LLMs可作为眼科图像解读的辅助工具,但对罕见病变、近视性病变和复杂血管疾病需谨慎解读。
- 结合GPT-4.5和Gemini 2.0 Pro的互补优势可能提升诊断支持效果。
- 需关注模型拒绝诊断的情况,避免依赖单一模型输出。
局限性
- 样本量有限(100张眼底图像),可能影响泛化性。
- 未报告模型在不同人群或图像质量下的表现。
- 评估基于特定验证图表,可能未全面反映临床实际场景。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND/AIMS: To evaluate and compare the diagnostic capabilities of advanced large language models (LLMs) in interpreting ophthalmological fundus images across diverse pathologies. METHODS: We evaluated eight leading multimodal LLMs (GPT-4.5, Claude 3.7 Sonnet, Grok-2, Deepseek Cognition V2, Qwen2 72B, Gemini 2.0 Pro, Llama 3 405B and Mixtral 8×22B) on their ability to interpret 100 fundus images representing various ophthalmological conditions. Performance was assessed using validated charts for diagnostic accuracy, specificity, sensitivity, consistency, relevance and explanation quality. RESULTS: GPT-4.5 achieved the highest overall diagnostic accuracy (65.0%), followed by Gemini 2.0 Pro (63.0%). All models showed varied performance across pathology categories, with rhegmatogenous pathologies being most accurately identified (Gemini 2.0 Pro: 81.3%, GPT-4.5: 75.0%) and myopic maculopathy (mean accuracy 21.8%) being particularly challenging. The remaining models performed significantly worse: Deepseek Cognition V2 (52.0%), Claude 3.7 Sonnet (52.0%), Qwen2 72B (49.0%), Llama 3 405B (48.0%), Grok-2 (47.0%) and Mixtral 8×22B (46.0%). Lower-performing models frequently declined to provide diagnoses, with refusal rates from 8.0% (Claude 3.7 Sonnet) to 19.0% (Mixtral 8×22B). CONCLUSION: Current LLMs show promising but limited capabilities in ophthalmological image interpretation. While performance on common conditions like retinal detachments and age-related macular degeneration is moderately good, significant challenges remain with rare conditions, myopic pathologies and complex vascular disorders. The competitive performance between GPT-4.5 and Gemini 2.0 Pro, with each excelling in different pathology categories, suggests that leveraging their complementary strengths might offer improved diagnostic support.
引用
- Carlà MM, Crincoli E, Catania F, et al. Advanced analysis of leading large language models for diagnostic accuracy in retinal imaging. Br J Ophthalmol. 2026. PMID: 41535097. doi:10.1136/bjo-2025-327634
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。