Automatic Hip Detection in Anteroposterior Pelvic Radiographs-A Labelless Practical Framework.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated Region of Interest (ROI) detection is crucial for medical image analysis, yet often lacks detailed reporting on parameters and confidence scores.
• Existing methods may not adequately address the variability and complexity of medical imaging datasets.

Purpose of the Study:

• To develop and validate a practical deep learning framework for robust automated detection of hip joints in anteroposterior pelvic radiographs (PXR).
• To address the common omission of essential details like model parameters, annotation rules, and confidence scores in medical image analysis studies.

Main Methods:

• A deep learning framework utilizing a single-shot multi-box detector with a customized head structure was developed.
• The model was trained and tested on a diverse dataset of 7399 pelvic radiographs from three distinct sources.
• Flexible loose-fitting labeling and heterogeneous data testing were employed to enhance robustness.

Main Results:

• The framework achieved high performance metrics on an independent testing set: average Intersection over Union (IoU) of 0.8115, average confidence of 0.9812, and average precision (AP50) of 0.9901.
• These results indicate that the detected hip regions accurately encompass the primary anatomical features.
• The study demonstrated the feasibility of training a reliable hip region detector for PXRs.

Conclusions:

• The proposed practical framework offers a robust solution for automated hip region detection in pelvic radiographs.
• The approach's flexibility in labeling and model design, coupled with testing on heterogeneous data, highlights its adaptability.
• This framework holds significant potential for enhancing various medical image analysis applications requiring precise ROI identification.