Learn2Segment — Medical Imaging Framework
Meta-learning framework for few-shot medical image segmentation, enabling U-Net-based models to generalise to new organs or modalities from as few as one labelled example.
Python 
PyTorch Learn2Segment is the research framework developed during my final-year internship at IRIT (Institut de Recherche en Informatique de Toulouse). It addresses a fundamental bottleneck in medical AI: annotated medical images are expensive and scarce, yet standard deep learning demands thousands of labelled examples.
The problem
Segmenting anatomical structures (organs, lesions, tumours) from MRI or CT scans requires pixel-level annotations made by radiologists — a slow, costly process. A model trained to segment the liver cannot trivially segment the spleen without starting over.
The approach
Learn2Segment applies meta-learning (learning-to-learn) to segmentation. The framework trains across many segmentation tasks so that when it encounters a new organ with only a handful of labelled slices, it can adapt quickly. The backbone is a U-Net architecture, pre-trained in a meta-learning configuration using MAML-style inner/outer loop optimisation.
Medical image handling uses MONAI (Medical Open Network for AI) for data loading, augmentation pipelines, and evaluation metrics (Dice, HD95), and SimpleITK for volumetric preprocessing (resampling, normalisation, orientation).
Context
This project directly informed my PhD research direction — the same challenge of data scarcity and domain-specific expertise applies to medieval Arabic manuscript recognition, just in a different modality.