3D Breast Cancer Segmentation in DCE-MRI Using Deep Learning With Weak Annotation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Acquiring large annotated datasets for deep learning in medical imaging is challenging.
• Weak annotation offers a time-efficient alternative for dataset creation.
• This study addresses the need for robust deep learning models in breast cancer imaging.

Purpose of the Study:

• To develop and evaluate a deep learning model for 3D breast cancer segmentation in DCE-MRI.
• To utilize weak annotation techniques for training the model.
• To achieve reliable segmentation performance for clinical applications.

Main Methods:

• A retrospective study involving 736 women with breast cancer.
• Weak annotation using bounding boxes, followed by manual correction, to create ground truth.
• Training a 3D U-Net transformer (UNETR) deep learning model on the annotated DCE-MRI dataset.
• Quantitative (Dice similarity, Spearman correlation) and qualitative (visual scoring) evaluation of segmentation results.

Main Results:

• The deep learning model achieved a median Dice similarity score of 0.75 for the whole breast and 0.89 for the region of interest (ROI).
• Volume correlation coefficients with ground truth were 0.82 (whole breast) and 0.86 (ROI).
• The mean visual score from reader evaluation was 3.4.

Conclusions:

• The developed deep learning model shows promising performance for 3D breast cancer segmentation in DCE-MRI using weak annotations.
• Weakly supervised learning can be effective for medical image segmentation tasks.
• This approach has the potential to improve the efficiency and accuracy of breast cancer diagnosis.