Can AI Detect What Is Not Injected? Evaluation of Lesion Detection in Virtual Contrast-Enhanced Breast MRI Using a Large-Scale AI Model Trained on GBCA-Enhanced Data.
Artificial intelligence (AI) can support lesion detection in gadolinium-based contrast agent-enhanced (GBCA-enhanced) breast MRI. However, its effectiveness on virtual contrast-enhanced (vCE) images remains unclear. This feasibility study evaluated the publicly available MAMA-MIA nnU-Net model trained on GBCA-enhanced data using an independent cohort of both GBCA-enhanced and vCE breast MRI.
This IRB-approved retrospective study included the publicly available nnU-Net model trained on n = 1506 MAMA-MIA breast MRI scans and a cohort of n = 2126 in-house 3T breast MRI scans. A generative adversarial network (Pix2Pix-GAN) was developed on n = 1870 of the in-house scans and used to generate vCE data on the remaining independent n = 256 in-house cases. The MAMA-MIA nnU-net was applied to both GBCA-enhanced (GBCA) and corresponding vCE images. Ground-truth segmentations of malignant lesions served to calculate the Dice score, Hausdorff distance, and lesion dimension differences.
The final test set comprised n = 250 cases (n = 69 malignant, n = 181 benign). Lesion detection rates were 91% (n = 63/n = 69; 95% confidence interval (CI): 82.3-96.0%) for GBCA and 84% (n = 58/n = 69; 95% CI: 73.7-90.9%) for vCE. Two lesions missed in GBCA were identified by vCE. The Hausdorff distances were similar (GBCA: 6.4 (IQR: 3.2-9.3; 95% CI: 5.2-7.8) mm; vCE: 6.7 (IQR: 3.9-9.7; 95% CI: 5.3-8.0) mm, p = 0.564). The Dice scores showed minor differences (GBCA: 0.829 (IQR: 0.723-0.900; 95% CI: 0.786-0.865) vs. vCE: 0.826 (IQR: 0.720-0.857; 95% CI: 0.770-0.836); p < 0.001). vCE images had slightly higher non-target tissue segmentation (median 6072 mm3 vs. 5754 mm3).
A GBCA-trained algorithm demonstrated some cross-domain transferability to vCE images, albeit with a reduced case-level sensitivity of 84% (95% CI: 73.7-90.9%) vs. 91% (95% CI: 82.3-96.0%). Based on these preliminary results, further research, including larger cohorts and more diverse datasets, is warranted.
This IRB-approved retrospective study included the publicly available nnU-Net model trained on n = 1506 MAMA-MIA breast MRI scans and a cohort of n = 2126 in-house 3T breast MRI scans. A generative adversarial network (Pix2Pix-GAN) was developed on n = 1870 of the in-house scans and used to generate vCE data on the remaining independent n = 256 in-house cases. The MAMA-MIA nnU-net was applied to both GBCA-enhanced (GBCA) and corresponding vCE images. Ground-truth segmentations of malignant lesions served to calculate the Dice score, Hausdorff distance, and lesion dimension differences.
The final test set comprised n = 250 cases (n = 69 malignant, n = 181 benign). Lesion detection rates were 91% (n = 63/n = 69; 95% confidence interval (CI): 82.3-96.0%) for GBCA and 84% (n = 58/n = 69; 95% CI: 73.7-90.9%) for vCE. Two lesions missed in GBCA were identified by vCE. The Hausdorff distances were similar (GBCA: 6.4 (IQR: 3.2-9.3; 95% CI: 5.2-7.8) mm; vCE: 6.7 (IQR: 3.9-9.7; 95% CI: 5.3-8.0) mm, p = 0.564). The Dice scores showed minor differences (GBCA: 0.829 (IQR: 0.723-0.900; 95% CI: 0.786-0.865) vs. vCE: 0.826 (IQR: 0.720-0.857; 95% CI: 0.770-0.836); p < 0.001). vCE images had slightly higher non-target tissue segmentation (median 6072 mm3 vs. 5754 mm3).
A GBCA-trained algorithm demonstrated some cross-domain transferability to vCE images, albeit with a reduced case-level sensitivity of 84% (95% CI: 73.7-90.9%) vs. 91% (95% CI: 82.3-96.0%). Based on these preliminary results, further research, including larger cohorts and more diverse datasets, is warranted.
Authors
Heidarikahkesh Heidarikahkesh, Schreiter Schreiter, George George, Nguyen Nguyen, Skwierawska Skwierawska, Brock Brock, Hadler Hadler, Uder Uder, Laun Laun, Ehring Ehring, Graber Graber, Döppmann Döppmann, Horishnyi Horishnyi, Kapsner Kapsner, Ohlmeyer Ohlmeyer, Liebert Liebert, Bickelhaupt Bickelhaupt
View on Pubmed