Poster R9 - Advancing Surgical Precision: Automated Landmark Detection for Intraoperative Femur Fracture Alignment Assessment

Introduction: This study introduces an innovative method to improve intraoperative alignment assessment during femur fracture surgeries. Current manual alignment estimation techniques based on biplanar X-ray images often result in postoperative malalignment, leading to a significant rate of reoperation. Our approach leverages a U-Net model with a MobileNetV2 backbone to automate landmark detection on X-ray images, using normal femur landmarks as a reference for precise adjustments during surgery. We curated a new dataset, RFX-1, focused on hip and knee landmark segmentation, and trained our neural network on this dataset. Our model achieved competitive precision, recall, and F1-score in accurately segmenting femur landmarks on over 550 X-ray images. Trained on annotated AP hip and knee X-ray images, our model facilitates alignment measurement, promising enhanced surgical outcomes by improving intraoperative alignment assessments and surgical precision.

Materials and

Methods: Using the developed dataset, RFX-1, a neural network underwent training to detect specific femur landmarks in fluoroscopy. This dataset comprises over 500 high-quality anonymized X-ray images, including upper and lower extremities. Ground truth segmentation masks, established by domain experts, formed the basis of our study. Knee and hip X-ray images were selected for annotation, categorized by X-ray view (frontal or lateral) and bone area (hip and knee). These images were processed using the Computer Vision Annotated Tool (Intel, USA), where segmentation masks were created by domain experts. Segmentation labels were based on anatomical landmarks crucial for calculating femoral length and rotation to enhance intraoperative alignment assessment. Leveraging this dataset, we fine-tuned a U-Net model with a MobileNetV2 backbone to extract vital features and autonomously segment specified femur landmarks with high precision. Evaluation of our model's performance was conducted using precision, recall, and F1-score metrics. 10-fold cross-validation was applied to validate the robustness of our model.

Results, Conclusions, and Discussions:

Results: This study annotated 560 anterior-posterior (AP) X-ray images of the hip and knee region with ground truth segmentation labels and anatomical landmarks following our specific methodology. Using this dataset, the U-Net model was trained, resulting in mean precision, recall, and F1-score of 0.86, 0.83, and 0.83, respectively, for simultaneous segmentation of the femur shaft, neck, epicondylar, and condylar axes. Additionally, we introduced a novel method utilizing these bone axes for alignment measurement.

Conclusions: By concentrating on femur landmark segmentation, particularly in the hip and knee areas crucial for assessing correct alignment during femur fracture surgery, we bridge a significant literature gap. Our study represents a targeted effort to accomplish this segmentation task, employing a well-established U-Net model with a MobileNetV2 backbone. This approach facilitates accurate and automated segmentation of femur landmarks, offering potential enhancements in intraoperative alignment assessment. Moreover, the segmented landmarks hold promise in guiding surgical robots for automatic alignment and aiding manual surgeries by providing essential measurement tools for assessing femur alignment intraoperatively.

Acknowledgements (Optional):