ACL Injury Classification Using Gemini 2.5 Pro: Evaluation of Prompting Strategies

Aims This study aimed to evaluate the performance of Google Gemini 2.5 Pro in classifying anterior cruciate ligament (ACL) status on knee MRI and compare three distinct prompting strategies. Methods A total of 150 proton-density fat-suppressed (PD-FS) MRI volumes, 50 healthy, 50 partially torn, and 50 completely torn were obtained from a publicly available dataset from the Clinical Hospital Centre Rijeka in Croatia (2006–2014). Multimodal inputs were provided via Gemini’s Python SDK. Three prompts were tested: a general series prompt, a technical-description prompt, and a region-of-interest (ROI)-focused prompt. Model outputs were compared to radiologist-assigned labels using accuracy, precision, recall, specificity, F1 score, confusion matrices, and mean inference time. Results Mean inference time was 2.1 ± 0.3 seconds. The ROI-focused prompt achieved the highest F1 score (0.31) and precision (0.32), while recall and specificity remained consistent across prompts. Confusion matrices indicated improved identification of complete tears with ROI prompting. Conclusions Prompt design significantly impacts the diagnostic performance of LLMs, with anatomically targeted (ROI-focused) prompts enhancing knee MRI interpretation.