Table. 2.

Performance of AI-CAD in mammography

Literature AI-CAD AUROC Sensitivity Specificity
Rodriguez-Ruiz et al. (2019)25) Deep CNN (Transpara 1.4.0; Screenpoint Medical) 0.84 75%–86% at radiologists’ specificity -
Kim et al. (2020)27) Deep CNN (Lunit INSIGHT MMG; Lunit Inc.) 0.94 88.87% 81.87%
Kim et al. (2018)28) Deep CNN 0.903–0.906 75.6%–76.1% 88.5%–90.2%
Ribli et al. (2018)29) Faster R-CNN 0.95 90% -
Becker et al. (2017)30) Deep ANN 0.81–0.85 59.8%–73.7% 69.6%–84.4%
Kooi et al. (2017)31) Deep CNN 0.929 - -
McKinney et al. (2020)32) Deep learning AI system 0.740 42%–100% 92%–97%
Lee et al. (2022)33) Deep CNN (Lunit INSIGHT MMG 1.1.1.0; Lunit Inc.) 0.915 - -
Rodríguez-Ruiz et al. (2019)34) Deep CNN (Transpara 1.3.0; Screenpoint Medical) With AI: 0.89 (vs. Without AI: 0.87) With AI: 86% (vs. Without AI: 83%) With AI: 79% (vs. Without AI: 77%)
Chen et al. (2023)35) Deep CNN (Lunit INSIGHT MMG 1.1.7.1; Lunit Inc.) 0.93 84%–91% 77%–89%
Salim et al. (2020)36) Three commercialized AI-CADs 0.920–0.956 67.0%–81.9% at radiologists’ specificity -
Hickman et al. (2022)37) ML algorithms used in 14 studies Pooled AUROC 0.89 Pooled sensitivity 75.4% Pooled specificity 90.6%
Yoon et al. (2023)38) AI algorithms used in 13 studies Pooled AUROC 0.87–0.89 Pooled sensitivity 75.8%–80.8% Pooled specificity 76.9%–95.6%
Akselrod-Ballin et al. (2019)40) A ML-DL model 0.91 87% 77.3%
Schaffter et al. (2020)42) A custom neural network 0.858–0.903 - 66.2%–81.2% at radiologists’ sensitivity
Pacilè et al. (2020)43) Deep CNN (MammoScreen V1; Therapixel) With AI: 0.797 (vs. Without AI: 0.769) With AI: 69.1% (vs. Without AI: 65.8%) With AI: 73.5% (vs. Without AI: 72.5%)

AI, artificial intelligence; AI-CAD, artificial intelligence-based computer-aided detection; AUROC, area under the receiver operating characteristics curve; CNN, convolutional neural network; ANN, artificial neural network; ML, machine learning; DL, deep learning.

Korean J Fam Pract 2023;13:196~210 https://doi.org/10.21215/kjfp.2023.13.4.196
© KJFP