Can AI Detect What Is Not Injected? Evaluation of Lesion Detection in Virtual Contrast-Enhanced Breast MRI Using a...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Artificial intelligence (AI) shows promise in supporting lesion detection in gadolinium-based contrast agent-enhanced (GBCA-enhanced) breast MRI.
• The effectiveness of AI models trained on GBCA-enhanced images for virtual contrast-enhanced (vCE) breast MRI remains largely unexplored.

Purpose of the Study:

• To evaluate the feasibility of using a publicly available AI model (MAMA-MIA nnU-Net) trained on GBCA-enhanced breast MRI for lesion detection on vCE images.
• To assess the cross-domain transferability of an AI model from GBCA-enhanced to vCE breast MRI datasets.

Main Methods:

• A retrospective study utilized an existing nnU-Net model trained on 1506 MAMA-MIA breast MRI scans.
• A generative adversarial network (Pix2Pix-GAN) was employed to create vCE images from an independent cohort of 256 in-house 3T breast MRI scans.
• The MAMA-MIA nnU-Net model was applied to both GBCA-enhanced and vCE images, with performance evaluated using Dice scores, Hausdorff distance, and lesion detection rates against ground-truth segmentations.

Main Results:

• The AI model achieved a lesion detection rate of 91% on GBCA-enhanced images and 84% on vCE images.
• Hausdorff distances were comparable between GBCA-enhanced (6.4 mm) and vCE (6.7 mm) images.
• Dice scores indicated minor differences (GBCA: 0.829 vs. vCE: 0.826), with slightly higher non-target tissue segmentation observed in vCE images.

Conclusions:

• The GBCA-trained AI algorithm demonstrated cross-domain transferability to vCE breast MRI images.
• While effective, the model exhibited a slight reduction in case-level sensitivity when applied to vCE images (84% vs. 91%).
• Further research with larger and more diverse datasets is recommended to validate these preliminary findings.