Development and Validation of a Deep-Learning-Based Algorithm for Detecting and Classifying Metallic Implants in...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Deep Learning Algorithms

Background:

• Metallic implants in CT scans can cause artifacts, complicating image interpretation.
• Accurate segmentation and classification of metallic objects are crucial for diagnostic accuracy and workflow optimization.

Purpose of the Study:

• To develop and validate a deep-learning-based algorithm (DLA) for segmenting and classifying metallic objects in abdominal and spinal CT topograms.
• To assess the performance and efficiency of the DLA compared to existing methods.

Main Methods:

• A U-net-like architecture was used for DLA training on annotated hip and spine implant topograms.
• The DLA was validated using internal and external datasets, with performance compared against two radiologists.
• Sensitivity, specificity, and intersection over union (IoU) were calculated for DLA performance evaluation.

Main Results:

• The DLA demonstrated high performance (>95% in internal validation) for implant segmentation and classification.
• External validation showed sensitivity and specificity >90% for spine and hip implants across various topograms.
• The DLA reduced reconstruction time by 27.4% on average and achieved an IoU >0.9 with radiologists.

Conclusions:

• A prototype DLA effectively detects metallic spine and hip implants on CT topograms.
• The DLA significantly improves the scan workflow, offering good performance for both spine and hip implants.
• This AI-driven approach enhances diagnostic capabilities in the presence of metallic artifacts.