The impact of common image quality issues in screening mammograms on commercial AI‑CAD models

Abstract

Artificial intelligence (AI)-based computer-aided detection (CAD) models for mammography have achieved performance levels comparable to or exceeding that of radiologists. This study aimed to investigate whether common image quality issues in mammography influence the performance of commercial AI-CAD models. A retrospective study was conducted using data from a previous observer study involving 80 digital mammography examinations (40:20:20 cancers:benign:normal) interpreted by 13 expert breast radiologists. Common image quality issues in acquisition and post-processing were simulated at three severity levels. Now, two commercial AI-CAD models were used to assess the same examinations. The area under the receiver operating characteristic curve (AUC) across quality levels was compared to those at the reference quality with a significance value of p=0.05. On reference-quality images, radiologists achieved an average AUC of 0.76 (95% CI: 0.68–0.84), AI A achieved 0.95 (95% CI: 0.90–1.00) and AI B 0.72 (95% CI: 0.60–0.83). Radiologists’ AUC remained stable across the three quality levels (0.77, 0.75, and 0.75, p=0.76, p=0.39, and p=0.63, respectively). In contrast, AI A showed significant decreases in AUC with poorer image quality (0.91, 0.83, and 0.78, p=0.27, p=0.02, and p=0.002, respectively), and AI B displayed a similar but nonsignificant trend (0.68, 0.61, and 0.65, p=0.47, p=0.06, and p=0.26, respectively). In conclusion, the performance of AI-CAD models can be affected by common image quality issues in mammography at levels of severity that do not affect radiologists. These findings highlight the importance of testing AI-CAD robustness conditions before clinical deployment and of quality control after deployment.

Ähnliche Arbeiten