A central challenge in biomedical image processing is the combination of low segmentation accuracy, small datasets, and low resolution. This work, conducted at Toronto Metropolitan University’s Maternal-Fetal Imaging Laboratory under Dr. Dafna Sussman, compared several approaches to brain metastasis segmentation.

Methods

We evaluated three families of models on a dataset of fewer than 100 3D MRI scans:

  1. U-Nets — encoder-decoder architectures with skip connections, widely used in biomedical segmentation
  2. Fully Convolutional Networks (FCNs) — end-to-end convolutional architectures adapted for dense pixel-wise prediction
  3. Gradient-Boosted Ensemble Models — classical machine learning approaches using hand-crafted features with XGBoost

Key Findings

Deep learning methods (particularly U-Net variants) significantly outperformed classical approaches on segmentation accuracy, even with limited training data. However, the gradient-boosted models offered faster training times and more interpretable feature importance rankings, which may be valuable in clinical settings where model transparency is required.

Data augmentation strategies—including elastic deformations, intensity shifts, and random rotations—proved critical for the deep learning models to generalize from such a small training set.

Collaborators