Complementary roles of convolutional neural networks and multimodal large language models (ChatGPT) in radiographic...

Accurate identification of implants in total knee arthroplasty (TKA) is essential for revision arthroplasty but is often complicated by incomplete documentation. While convolutional neural networks (CNNs) achieve high accuracy in implant recognition, the role of multimodal large language models (LLMs), such as ChatGPT remains unclear.

This study combined a retrospective radiographic analysis of 80 bicompartmental TKA cases with a structured comparison of CNN-based classification and LLM-based interpretive reasoning. The analysis included four implant systems. Convolutional neural networks served as benchmark references, while a multimodal large language model (ChatGPT) was evaluated for implant identification, reasoning, and confidence estimation.

Radiographs in anteroposterior projection showed higher recognition performance than lateral views across all systems. The combined use of both projections improved reliability, indicating dependence on complementary morphological features. Convolutional neural network-based approaches achieved accuracies exceeding 99%. In contrast, the evaluated large language model provided contextual interpretation and implant suggestions but lacked measurable classification accuracy. Confidence levels were moderate (median 0.63-0.70). Interrater agreement was moderate to substantial (κ = 0.52-0.74).

Convolutional neural network-based algorithms remain the gold standard for implant classification. Multimodal large language models provide complementary strengths in contextual reasoning. Hybrid approaches may improve clinical workflows in revision arthroplasty.