A unified deep learning framework for multi-vessel classification, segmentation and phenomapping of phase-contrast MRI validated on a multi-site single ventricle patient cohort

This repository presents a deep learning pipeline for fully automated segmentation and downstream analysis of velocity-encoded phase-contrast magnetic resonance (PCMR) images. The primary application is within Fontan circulation patients, leveraging data from the FORCE Registry, but the methods are generalizable to other cardiovascular conditions.

Tina Yao, Nicole St. Clair et al. MultiFlow: A unified deep learning framework for multi-vessel classification, segmentation and clustering of phase-contrast MRI validated on a multi-site single ventricle patient cohort [preprint] [1]

Motivation

Time-varying PCMR signals offer rich insights into cardiovascular physiology—beyond what static metrics like stroke volume can provide. However, large registries like FORCE (with >5,000 Fontan exams) primarily store raw PCMR data and reported metrics, with no time-resolved flow curves extracted.

Manual segmentation of these datasets is infeasible (>1,000 hours). MultiFlowSeg was designed to automate segmentation and extract time-varying flow curves across five key vessels:

• Aorta (Ao)
• Left/Right Pulmonary Arteries (LPA/RPA)
• Superior/Inferior Vena Cava (SVC/IVC)

These data are then analyzed using Deep Temporal Clustering (DTC) to uncover patient subtypes and their association with clinical outcomes (e.g., liver disease, transplant/mortality).

Project Goals

1. Develop a unified DL model capable of classifying and segmenting multiple vessel types from PCMR slices (MultiFlowSeg).
2. Build a scalable, automated pipeline for segmenting the entire FORCE registry of >5000 CMR exams (Pipeline).
3. Extract time-varying flow curves and perform unsupervised phenomapping using Deep Temporal Clustering (MultiFlowDTC).
4. Correlate clusters with outcomes like liver disease, and transplant/mortality.

MultiFlowSeg Architecture

Figure 1: Model architecture for multi-vessel segmentation and classification.

Segmentation & Derived Flow Curves

Figure 2: Example of segmentation results and derived time-varying flow curves for key vessels.

MultiFlowDTC Architecture

Figure 3: Model architecture for unsupervised clustering of flow curves.

Clusters and Phenomapping

Figure 4: Clustered flow curves and associated phenomapping outcomes.