Automated 3T DCE-MRI Breast Lesion Segmentation: A Two-Stage V-Net Deep Learning Approach for Enhanced Clinical Efficiency
Abstract
Purpose
In medical imaging, the accurate segmentation of lesions is a fundamental step for monitoring tumor progression and extracting quantitative radiomic features for lesion characterization. Traditionally, this task has relied on manual input, which is notoriously time-consuming and heavily influenced by operator variability, often impacting the consistency of subsequent analyses. To address these challenges, this study developed an automated deep learning (DL) model based on the V-Net architecture specifically for malignant breast lesion segmentation on 3T DCE-MRI.
Methods
This retrospective study involved 131 women undergoing DCE-MRI prior to neoadjuvant chemotherapy between 2016 and 2021. A two-stage V-Net approach was implemented. In the first stage, an initial model was trained to generate a raw breast mask with ample borders, effectively restricting the search area for subsequent detection. In the second stage, the breast region was divided into smaller 3D patches to capture fine-grained features, where a second V-Net model identified the specific lesions. Fine-tuning was then applied to these patches to establish optimal binarization thresholds and minimize false positives. The Dice Similarity Coefficient (DSC) was used to evaluate the performance compared to two expert readers.
Results
The study evaluated various parameters, finding that utilizing 100% background patches combined with a 0.6 binarization threshold yielded the most accurate results. This configuration achieved a median Dice score of 0.77 for the training set and 0.71 for the test set when compared to manual segmentation. Notably, the automated system was approximately five times faster than manual processes, significantly enhancing workflow efficiency.
Conclusion
The proposed model successfully reduces human operator variability and provides a highly efficient alternative to manual methods. However, human supervision remains essential to address false positives and fine-tune results. While current outcomes are satisfactory, further improvements could be achieved through dataset augmentation, architectural modifications, and hardware upgrades.