High-Performance Open-Source AI for Breast Cancer Detection and Localization in MRI.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Breast cancer diagnosis relies heavily on Magnetic Resonance Imaging (MRI).
• Accurate detection and localization are critical for effective treatment planning.
• Developing advanced computational tools can aid radiologists in improving diagnostic accuracy and efficiency.

Purpose of the Study:

• To develop and validate an open-source deep learning model for breast cancer detection and localization using MRI.
• To assess the model's performance on a large, diverse dataset and evaluate its generalizability across different imaging planes and clinical sites.

Main Methods:

• A retrospective study utilizing the largest breast MRI dataset to date.
• Training a deep learning model on over 30,000 sagittal MRI examinations.
• Validating the model on sagittal and axial MRI data from primary and secondary clinical sites.

Main Results:

• The model achieved an area under the receiver operating characteristic curve (AUC) of 0.95 for cancer detection on primary site sagittal data.
• Sensitivity was 83% at 90% specificity, comparable to radiologist performance.
• The model demonstrated strong generalizability with AUCs of 0.92 on axial data from both primary and secondary sites, and accurately localized tumors in over 87% of cases across datasets.

Conclusions:

• The developed deep learning model exhibits state-of-the-art performance for breast cancer detection and localization on MRI.
• The open-source nature of the code and weights encourages further research, validation, and clinical integration.
• This AI tool has the potential to significantly enhance computer-aided diagnosis (CAD) in breast MRI reviews.