A Pipelined Tracer-Aware Approach for Lesion Segmentation in Breast DCE-MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep Learning (DL) shows promise in medical imaging, particularly for Dynamic Contrast-Enhanced Magnetic-Resonance Imaging (DCE-MRI) in breast cancer analysis.
• Naive application of DL to DCE-MRI is limited due to the complexity of multimodal 4D image data and the need for specialized preprocessing.

Purpose of the Study:

• To develop and evaluate a comprehensive pipelined approach for enhancing DL-based lesion segmentation in breast DCE-MRI.
• To address the unique challenges of DCE-MRI data, including contrast agent dynamics, non-breast tissues, and patient motion.

Main Methods:

• A novel pipeline incorporating breast-masking, Three-Time-Points (3TP) slice selection, motion correction, a modified U-Net architecture, and an "Eras/Epochs" training strategy.
• The method was designed to handle unbalanced datasets and incorporate robust data augmentation.

Main Results:

• The proposed pipelined approach significantly outperformed existing literature methods in lesion segmentation accuracy.
• Demonstrated a substantial improvement of +9.13% over the previous solution, indicating superior performance.
• Showcased enhanced generalization ability on unseen data.

Conclusions:

• The developed pipelined strategy effectively leverages the specific characteristics of breast DCE-MRI data for improved lesion segmentation.
• This tailored approach offers a more effective solution than naive DL applications for this complex imaging modality.
• The findings suggest a promising direction for advancing AI-driven breast cancer analysis using DCE-MRI.