🧠 Segmenting Infant Brains Across Magnetic Fields

Domain Randomization and Annotation Curation in Ultra-Low-Field MRI

This repository contains the source code, model weights, and image quality annotations used in our MICCAI 2025 LISA Challenge paper:
“Segmenting infant brains across magnetic fields: Domain randomization and annotation curation in ultra-low-field MRI” ([DOI]).

🏆 Our method ranked 1st in hippocampus segmentation and 3rd in basal ganglia segmentation at the MICCAI 2025 LISA Challenge Challenge link.

---

🩻 Quality Annotations

To ensure high-quality training data, we manually curated image quality annotations by identifying and excluding cases with severe segmentation misregistration.

Each LISA 2025 training case was visually inspected and labeled as either:

• good – proper alignment between segmentation and image contrast (especially along ventricular boundaries)
• bad – visible misalignment or deformation

The resulting annotations are provided in annotations file link.
These can be used to filter out problematic cases before training.

---

⚙️ Model Weights

Pre-trained model weights from all experiments are available at OneDrive link.

They can be used directly with the
inference.py
and
inference_lisa.yaml
files for inference.

🔧 Usage

The inference script expects:

• A BIDS-formatted dataset
• A split file (CSV) with:
  • participant_id – BIDS subject IDs
  • splits – split names

Only subjects matching the split_file value defined in the YAML configuration will be used for inference.

Example config entry:

split_file: /path/to/splits.csv