Large language models show potential but lag behind expert veterinary radiologist in calf septic arthritis detection

A study finds that while ChatGPT-5 and other large language models perform well in detecting obvious calf septic arthritis, they still fall short of expert human readers, especially on subtle lesions, highlighting their role as supportive tools rather than standalone diagnostic solutions.

General-purpose large language models can support structured radiographic review in calves, but their performance remains uneven and below that of the strongest novice reader. In a reader study of presumptive septic carpal arthritis, ChatGPT-5 emerged as the best of three models, yet it still lagged behind the most accurate human evaluator and was less reliable on subtle findings and composite scores.

The study used 50 single-view radiographs, a Constant-based ordinal rubric and a short clinical prompt to test two novice veterinary surgeons against ChatGPT-5, Gemini-2.5 Pro and Claude Sonnet-4. According to the authors, the best human reader achieved the highest overall concordance with an expert consensus reference, while ChatGPT-5 was the strongest model and sometimes matched or exceeded the weaker novice on conspicuous osseous changes.

Performance was notably better for more obvious lesions, such as bone proliferation, subchondral lysis and sclerosis, than for softer features like swelling and joint-space change. This mirrors earlier veterinary work showing that septic arthritis in calves is associated with articular widening, surface irregularity and subchondral lysis, and that structured scoring can improve consistency.

The authors caution that their results reflect agreement with expert-derived scores, not definitive pathology. Still, the findings matter for bovine practice: multimodal LLMs may have value as adjuncts for training and first-pass review, but not as stand-alone readers.

Source: Noah Wire Services