Automated Detection of Hip Replacements and Fiducial Markers in Pelvic CT Scans: A Comparative Study of Rule-Based vs Deep Learning Approaches
This paper presents a comprehensive comparative study of automated detection of hip replacements and fiducial markers in pelvic computed tomography (CT) scans, evaluating rule-based anatomically-constrained methods against modern deep learning techniques. We develop and independently evaluate two distinct approaches: (1) a novel rule-based system utilizing metallic implant detection, fiducial marker identification, and anatomical constraint analysis specifically designed for hip replacement detection, and (2) a convolutional neural network (CNN) enhanced with Gradient-weighted Class Activation Mapping (Grad-CAM) for hip prosthesis and fiducial detection. Our evaluation on a comprehensive dataset of 444 pelvic CT cases demonstrates that the rule-based approach achieves superior performance (sensitivity: 90.4%, specificity: 88.2%, accuracy: 89.0%) while providing interpretability and clinical reasoning. The CNN achieves 70.8% ± 5.4% accuracy with GradCAM visualizations enabling spatial understanding of predictions for hip prosthesis and fiducial detection. This comparative analysis reveals that specialized rule-based methods can significantly outperform deep learning approaches while offering the interpretability essential for clinical adoption in medical imaging applications focused on data curation and quality control.
