Diagnostic Performance of AI-CAD Digital Mammography for Breast Cancer: Experience from Siriraj Breast Imaging Center

Abstract

Objective To evaluate the diagnostic performance of radiologists with varying breast imaging experience when interpreting digital mammograms with and without artificial intelligence (AI). This work represents the initial phase of an AI development program at Siriraj Hospital, aiming toward broader integration of AI into breast cancer detection and clinical practice in Thailand. Materials and Methods: In this retrospective study, six radiologists independently reviewed 86 digital mammograms— 40 confirmed cancer cases and 46 normal cases (including 28 false positives and 18 true negatives) — collected between 2018 and 2019 at the Siriraj Breast Imaging Center. Each radiologist interpreted all cases twice: unaided and AI-assisted, with a two-week washout period to minimize recall bias. Diagnostic performance metrics included sensitivity, specificity, false positive/negative rates, and reading time. Results: With AI assistance, sensitivity increased in five of six readers, with mean sensitivity rising from 56.1% to 77.5%, although this difference did not reach statistical significance. Changes in specificity were variable across readers, with a statistically significant improvement observed in one reader (52.2% to 78.3%, P < 0.05). Mean reading time decreased from 32.9 seconds to 21.0 seconds per case with AI assistance (P < 0.01), with reductions observed for both cancer cases and normal cases. Conclusion: In this pilot study, AI assistance was associated with trends toward improved diagnostic performance and reduced reading time, with statistically significant improvement observed in only a subset of readers. These preliminary findings require confirmation in larger, adequately powered multi-reader multi-case (MRMC) studies.