Software

Libraries I maintain, the code behind each paper, and fixes I have sent upstream, on GitHub.

Libraries

medrs
Medical imaging I/O and processing for deep learning pipelines. Memory-mapped NIfTI reading, crop-first loading of sub-volumes, and lazy pipelines that combine intensity operations in one pass and spatial operations in one interpolation. The repository is named med-rs and the package is published as medrs on crates.io and PyPI.
Rust, Python
gaze
Agentic framework for medical vision-language models: viewer-level image tools, multi-turn tool use, and literature retrieval from PubMed and Open-i. Published on PyPI as gaze-vlm.
paper: GAZE
Python
medtokenizers
Tokenizers for 2D and 3D medical images behind one encoder-decoder interface: VQ, LFQ, FSQ, residual FSQ, VAE and AE heads. Published on PyPI.
paper: Tokenizer-generator coupling
Python, PyTorch
medlatents
Generative models over tokenized medical images: autoregressive transformers, MaskGIT, discrete diffusion (D3PM), discrete flow matching, Bayesian flow networks, latent diffusion and rectified flow. Published on PyPI.
paper: Tokenizer-generator coupling
Python, PyTorch

tokenizer-generator-coupling
Reproduction code, result data and figure scripts for the tokenizer-generator coupling study.
paper: Tokenizer-generator coupling
Python, NeurIPS 2026
SynthStroke
Code and pretrained weights for stroke lesion segmentation trained entirely on synthetic data, with no sequence-specific training images. Also packaged as an SPM12 toolbox for use without a Python environment.
paper: SynthStroke
Python, MELBA 2025
qsynth
Physics-constrained synthetic qMRI generation for domain-agnostic lesion segmentation. Implements qATLAS, which estimates qMRI maps from MPRAGE, and qSynth, which synthesises them directly from tissue labels.
paper: qSynth
Python, MICCAI 2025
contrast-squared
Sequence-invariant contrastive pre-training for 3D MRI. A single encoder transfers across segmentation and denoising tasks and across sites. Weights are on Hugging Face.
paper: Sequence-invariant SSL
Python, SASHIMI 2025
MDUNet
All-convolutional transformer block for 3D brain lesion segmentation, giving transformer-like long-range modelling at CNN parameter cost.
paper: LKA
Python, NeurIPS 2023 workshop
RectAngle
Segmentation and classification for trans-rectal B-mode ultrasound, including the pre-screening classifier that suppresses false positives on frames with no region of interest.
paper: Intraoperative ultrasound
Python, MICCAI 2021 workshop

Project-MONAI/MONAI
Fixed dtype conversion and Spacing inverse-transform bugs affecting MONAI Label.
balbasty/nitorch
Added Google Colab support.
balbasty/cornucopia
Fixed a randint call in the geometric augmentation layers.