Performance of multimodal large language models for the detection and characterization of bone lesions on radiographs.

Authors

Abstract

Multimodal large language models (LLMs) offer emerging capabilities in medical image interpretation; however, their efficacy in orthopedic oncology remains unverified. This study aimed to evaluate and benchmark the performance of five contemporary LLMs-ChatGPT 5.2, Gemini 3 Flash, MedGemma 4B, Claude Sonnet 4.6, and DeepSeek-VL2-in detecting and characterizing bone lesions on plain radiographs without task-specific fine-tuning. A retrospective analysis was conducted using 3,746 anonymized images from the Bone Tumor X-ray Radiograph Dataset (BTXRD), comprising normal, benign, and malignant cases. Reference standard annotations were provided directly by the BTXRD dataset. Models were evaluated on two tasks: lesion detection and lesion characterization. Diagnostic performance metrics, including accuracy, precision, sensitivity, specificity, and Cohen's kappa, were calculated and compared with reference-standard annotations. ChatGPT 5.2 demonstrated the highest overall accuracy (0.803) and specificity among the models (0.916) for lesion detection, although its sensitivity (0.689) was comparatively low. MedGemma 4B showed relatively low performance, with an overall accuracy of 0.677. Claude Sonnet 4.6 and Gemini 3 Flash had the highest sensitivities among the models (0.991 and 0.972, respectively) but low specificities (0.038 and 0.201, respectively), resulting in excessive false positives. In the characterization task, ChatGPT 5.2 consistently achieved the highest performance among the models, with an accuracy of 0.758 and a weighted F1 score of 0.745. DeepSeek-VL2 achieved high specificity but very low sensitivity for malignancy (0.714 and 0.022, respectively). Gemini 3 Flash provided high sensitivity for malignancy (0.711) but low overall accuracy. Multimodal LLMs demonstrated heterogeneous performance in the evaluation of bone lesions on plain radiographs, with substantial differences across models and tasks. Although some models achieved high accuracy in lesion detection and overall classification, performance was inconsistent across tasks, particularly in identifying malignant lesions and balancing sensitivity and specificity. These findings suggest that, despite their potential, current multimodal LLMs are not yet sufficiently reliable for diagnostic use in orthopedic oncology and should be considered investigational until further development and validation. Multimodal LLMs currently lack the diagnostic reliability required for bone lesion assessment, often exhibiting excessive false positives or failing to detect malignancy. Although generalist models show promise, expert radiologist oversight remains essential to ensure patient safety and oncologic accuracy in musculoskeletal imaging.

Tags

Topics