Cross-site Validation of AI Segmentation and Harmonization in Breast MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Oncology

Background:

• Accurate segmentation of breast cancer in MRI is crucial for diagnosis and treatment planning.
• Variability in segmentation performance across different clinical sites poses a challenge for AI model generalizability.

Purpose of the Study:

• To validate the cross-site performance of an automated breast cancer segmentation model using a 3D U-Net architecture.
• To compare the AI model's segmentation accuracy against that of experienced radiologists.

Main Methods:

• A 3D U-Net model was trained on a large dataset of dynamic contrast-enhanced axial MRIs from Site 1.
• Performance was evaluated on test data from three clinical sites and common public data, with radiologist segmentations serving as ground truth.
• A supervised harmonization technique with an affine input layer was employed for Site 3 data.

Main Results:

• The AI model achieved segmentation performance comparable to radiologists across Sites 1, 2, and common public data (median Dice scores ranging from 0.86 to 0.93).
• For Site 3, fine-tuning the model resulted in performance on par with radiologists (Dice score 0.88 vs. 0.89).
• The AI model numerically outperformed 11 out of 12 individual radiologists on common test data.

Conclusions:

• A deep learning model, enhanced with a novel supervised harmonization technique, demonstrates radiologist-level performance in cross-site MRI breast cancer segmentation.
• The study highlights the potential of reproducible AI tools to aid in radiological assessments.
• Publicly released code and weights aim to foster further advancements in AI for radiology.