2621 - A Patch-Based Deep Learning Approach for Brain Metastases Segmentation: Balanced Sampling Strategy for Class Imbalance

Purpose/Objective(s): We hypothesized that an automated patch-based deep learning approach using a U-Net architecture would provide robust segmentation of Gross Tumor Volume (GTV) in T1-enhanced MRI scans of brain metastases. Our primary endpoint was the Dice Similarity Coefficient (DSC); secondary endpoints included a reduction in manual delineation time and inter-observer variability.

Materials/Methods: We collected T1-enhanced MRI images and corresponding GTV contours of 371 brain metastases from patients treated with the HyperArc technique at one institution, combined with open-source data from the Mathematical Oncology Laboratory. DICOM-RT structure files and all other data sources were converted to NIfTI format for standardized processing. Data preprocessing included standardized resolution (512×512), 3D resampling, and automated mask generation. Various data augmentation techniques were employed, including random rotations, scaling, elastic deformations, intensity adjustments, and Gaussian noise addition, to enhance model robustness, in the model training procedure. To address the class imbalance, we implemented a hybrid patch sampling strategy: uniform grid sampling for background regions and random-shift sampling near tumor regions, both with patch size (64×64×64). This approach generated a total of 146,759 training patches, ensuring comprehensive coverage while maintaining focus on tumor regions. A 3D U-Net-based architecture was trained and validated using 5-fold cross-validation, with 80% of patient data used for training and validation, and the remainder reserved for testing.

Results: Our model demonstrated consistent performance across all five cross-validation folds (training DSC: 0.9149-0.9289; validation DSC: 0.8529-0.9157), indicating successful management of class imbalance. The automated segmentation process achieved an average prediction time of 2 min 10 s (±15 s), significantly reducing manual delineation workload while maintaining accurate tumor boundary delineations.

Conclusion: Our standardized, patch-based deep learning pipeline using 3D U-Net-based architecture achieves reliable GTV segmentation of brain metastases in T1-enhanced MRI. Future research will focus on expanding our dataset through multi-center collaboration and conducting comprehensive multi-center validation studies.