Influence of content-based image retrieval on the accuracy and inter-reader agreement of usual interstitial pneumonia CT pattern classification

This study evaluated the impact of content-based image retrieval (CBIR) on the diagnostic accuracy and inter-reader agreement in classifying usual interstitial pneumonia (UIP) patterns on HRCT. A total of 321 CT images from patients with interstitial lung disease were interpreted by three radiologists first without CBIR and then with CBIR support. The CBIR system automatically retrieved similar cases based on visual features. Results showed that the use of CBIR significantly improved classification accuracy and inter-reader agreement (kappa values). The most notable improvement was seen in cases with intermediate likelihood of UIP, where the CBIR support helped reduce diagnostic ambiguity. This study demonstrates that AI-driven image retrieval systems can serve as effective decision support tools in radiology.