6 October 2022 in Journal of Orthopaedic Translation
Loosening is the leading cause of total knee arthroplasty (TKA) revision. This is a heavy burden toward the healthcare system owing to the difficulty in diagnosis and complications occurring from the delay management. Based on automatic analytical model building, machine learning, may potentially help to automatically recognize the risk of loosening based on radiographs alone. The aim of this study was to build an image-based machine-learning model for detecting TKA loosening. Image-based machine-learning model was developed based on ImageNet, Xception model and a TKA patient X-ray image dataset. Based on a dataset with TKA patient clinical parameters, another system was then created for developing the clinical-information-based machine learning model with random forest classifier. In addition, the Xception Model was pre-trained on the ImageNet database with python and TensorFlow deep learning library for the prediction of loosening. Class activation maps were also used to interpret the prediction decision made by model. Two senior orthopaedic specialists were invited to assess loosening from X-ray images for 3 attempts in setting up comparison benchmark. In the image-based machine learning loosening model, the precision rate and recall rate were 0.92 and 0.96, respectively. While for the accuracy rate, 96.3% for visualization classification was observed. However, the addition of clinical-information-based model, with precision rate of 0.71 and recall rate of 0.20, did not further showed improvement on the accuracy. Moreover, as class activation maps showed corresponding signals over bone-implant interface that is loosened radiographically, this confirms that the current model utilized a similar image recognition pattern as that of inspection by clinical specialists. The image-based machine learning model developed demonstrated high accuracy and predictability of knee arthroplasty loosening. And the class activation heatmap matched well with the radiographic features used clinically to detect loosening, which highlighting its potential role in assisting clinicians in their daily practice. However, addition of clinical-information-based machine-learning model did not offer further improvement in detection. As far as we know, this is the first report of pure image-based machine learning model with high detection accuracy. Importantly, this is also the first model to show relevant class activation heatmap corresponding to loosening location. The finding in this study indicated image-based machine learning model can detect knee arthroplasty loosening with high accuracy and predictability, which the class activation heatmap can potentially assist surgeons to identify the sites of loosening.
20 January 2023 in International Journal of Mechanical Sciences
27 August 2026 in SSRN Electronic Journal
Background and Objectives: Natural language interfaces (NLIs) have the potential to improve interaction with computer-assisted orthopedic surgery systems by reducing dependence on conventional mouse- and keyboard-based workflows. However, translating unstructured surgeon instructions into reliable and executable 3D surgical manipulations remains challenging due to linguistic variability and the strict precision requirements of surgical planning. This study presents a safety-constrained framework for evaluating Large Language Model (LLM)-based tool calling in Reverse Shoulder Arthroplasty (RSA) planning.Methods: We benchmarked Llama-3.1-8B, Qwen2.5-7B, and Mistral-v0.3-7B under zero-shot, few-shot, Retrieval-Augmented Generation (RAG), and Low-Rank Adaptation (LoRA) fine-tuning configurations. A synthetic benchmark containing 2,788 RSA planning commands with controlled linguistic perturbations was developed to evaluate structured JSON-based tool invocation. To investigate practical robustness, the optimized models were additionally evaluated using 84 surgeon-generated voice commands transcribed by OpenAI Whisper.Results: Results demonstrate that domain-specific adaptation substantially improves tool-calling reliability. The LoRA-adapted Qwen2.5-7B model achieved an Exact Match accuracy of 97.94% on the synthetic benchmark, significantly outperforming zero-shot baselines (p<0.001).Conclusions: The proposed safety-constrained architecture demonstrates the technical feasibility of using LLMs as an interaction layer between surgeon intent and computer-assisted surgical planning systems. Importantly, by routing probabilistic LLM outputs through deterministic schema validation and anatomical constraint checking before 3D rendering, the system ensures high precision and safety required for digital surgery.