liamchalcroft/midl-2025-seq-inv

SimCLR-MRI Pre-trained Encoder (SeqInv)

This repository contains a pre-trained 3D CNN encoder for MRI analysis. The model was trained using contrastive learning (SimCLR) with explicit sequence invariance enforced through paired multi-contrast images.

Model Description

The encoder is a 3D CNN with 5 convolutional blocks (64, 128, 256, 512, 768 channels), outputting 768-dimensional features. This SeqInv variant was trained on paired sequences generated through Bloch simulations, explicitly enforcing sequence invariance in the learned representations.

Training Procedure

• Pre-training Data: 51 qMRI datasets (22 healthy, 29 stroke subjects)
• Training Strategy: Paired sequence views + standard augmentations
• Input: 3D MRI volumes (96×96×96)
• Output: 768-dimensional sequence-invariant feature vectors

Intended Uses

This encoder is particularly suited for:

• Sequence-agnostic analysis tasks
• Multi-sequence registration
• Cross-sequence synthesis
• Tasks requiring sequence-invariant features