Deep-learning-based 3D content-based image retrieval system on chest HRCT: Performance assessment for interstitial...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Diffuse parenchymal lung diseases present diverse conditions and CT imaging findings.
• Distinguishing interstitial lung diseases (ILDs) and identifying usual interstitial pneumonia (UIP) is diagnostically challenging.
• A 3D content-based image retrieval (CBIR) system was developed to address these challenges.

Purpose of the Study:

• To develop and evaluate a 3D content-based image retrieval (CBIR) system for analyzing high-resolution computed tomography (HRCT) images.
• To assess the clinical usefulness of the CBIR system in differentiating ILDs and identifying UIP.
• To investigate the system's performance in image similarity and label concordance.

Main Methods:

• A deep learning-based prototype system was created to analyze and register thin-slice whole lung HRCT images.
• Search performance was evaluated using a database of 2058 cases with a 5-point visual similarity score.
• Clinical usefulness was assessed by evaluating label concordance (ILD/non-ILD, with/without UIP) on 301 cases across 57 diseases.

Main Results:

• The CBIR system achieved a mean visual similarity score of 4.37 ± 0.83 for top retrieved cases.
• Label concordance for ILD was high at 0.94 ± 0.15, and for non-ILD at 0.64 ± 0.31.
• Label concordance for cases with UIP was 0.86 ± 0.17, and for cases without UIP was 0.83 ± 0.24.

Conclusions:

• The developed CBIR system demonstrates high accuracy in identifying cases with and without usual interstitial pneumonia (UIP).
• The system shows significant potential to assist radiologists in the differentiation of UIP in clinical settings.
• This AI-driven approach offers valuable support for diagnosing complex interstitial lung diseases.