Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Deep Learning Pipeline for Assessing Ventricular Volumes in Single Ventricle Physiology (FORCE Registry)

Link

https://github.com/Ti-Yao/Single-Ventricle-Segmentation-Pipeline

Indexing

Keywords: Fontan, single ventricle, cardiac MRI, cine, ventricular volume, segmentation, U-Net 3+, myocardial mass, ejection fraction, FORCE registry

Content: CA, MR, PD

RadLex: RID4970, RID10312, RID10577, RID6257, RID1571

Author(s)

Tina Yao

Nicole St. Clair

Gabriel F. Miller

Adam L. Dorfman

Mark A. Fogel

Sunil Ghelani

Rajesh Krishnamurthy

Christopher Z. Lam

Michael Quail

Joshua D. Robinson

David Schidlow

Timothy C. Slesnick

Justin Weigand

Jennifer A. Steeden

Rahul H. Rathod

Vivek Muthurangu

Organization(s)

University College London

Boston Children's Hospital

University of Michigan

Children's Hospital of Philadelphia

Nationwide Children's Hospital

The Hospital for Sick Children, Toronto

Ann and Robert H. Lurie Children's Hospital of Chicago

Emory University School of Medicine

Texas Children's Hospital

Version

1.0

Contact

Corresponding author: Vivek Muthurangu; email: ku.ca.lcu@ugnaruhtum.v

Funding

Supported by a grant from the Additional Ventures Foundation. T.Y. supported by UKRI Centre for Doctoral Training in AI-enabled Healthcare Systems. Additional author disclosures list UKRI Future Leaders Fellowship (J.A.S.), British Heart Foundation support to V.M., and other grants as noted.

Ethical review

Multicenter retrospective study approved by Boston Children's Hospital IRB (IRB-P00028482) with waivers of consent; contributing institutions relied on BCH IRB or obtained local IRB/ethics approval with waivers of consent.

Date

Updated: 2024-01-01

Published: 2023-11-15

Created: 2023-11-15

References

[1] Yao T, St. Clair N, Miller GF, Dorfman AL, Fogel MA, Ghelani S, Krishnamurthy R, Lam CZ, Quail M, Robinson JD, Schidlow D, Slesnick TC, Weigand J, Steeden JA, Rathod RH, Muthurangu V. "A Deep Learning Pipeline for Assessing Ventricular Volumes from a Cardiac MRI Registry of Patients with Single Ventricle Physiology". Radiology: Artificial Intelligence. 2024 Jan;6(1):e230132. Published online 2023 Nov 15.. 2023-11-15. doi:10.1148/ryai.230132. PMID: 38166332. PMCID: PMC10831511.

Model

Architecture

End-to-end pipeline comprising: (1) CNN classifier for SAX identification; (2) U-Net 3+ for heart localization/cropping with full-scale skip connections and deep supervision; (3) U-Net 3+ for ventricular segmentation (blood pool, myocardium, background) with Tversky loss.

Availability

Open-source code available: https://github.com/Ti-Yao/Single-Ventricle-Segmentation-Pipeline

Clinical benefit

Automated, standardized ventricular segmentation and quantification (EDV, ESV, stroke volume, ejection fraction, myocardial mass) for patients with single ventricle physiology, enabling rapid large-scale analysis of registry data.

Clinical workflow phase

Post-acquisition image analysis and quantification; core lab standardization for research registries.

Decision threshold

SAX classifier threshold P_sax > 0.5 per image; SAX stack chosen by highest P_sax/mean P_sax. Dilated ventricle classification for analysis defined as EDV ≥ 156 mL/BSA^1.3.

Degree of automation

Fully automated end-to-end pipeline requiring no human input for processing; occasional user adjustments needed in a minority of cases.

Indications for use

Automated segmentation and ventricular quantification in cardiac MRI cine studies of patients with functionally single ventricle physiology (Fontan circulation) in a registry/research environment across multi-center, multi-vendor data.

Input

Cardiac MRI DICOM studies; cine stacks automatically extracted with focus on short-axis (SAX) cines; 2D cine frames across all slices and phases.

Instructions

Run pipeline on full CMR study; stages: (1) cine stack extraction using DICOM headers; (2) SAX identification via CNN; (3) heart localization and cropping via U-Net 3+ using ED frames to generate bounding box then crop all frames; (4) ventricular segmentation via U-Net 3+ for all slices/phases; volumes and EF derived from volume-time curves with ED/ES detected automatically.

Limitations

Performance reduced with suboptimal image quality and at 3T compared to 1.5T; trained/evaluated on single ventricle physiology, not validated on normal biventricular controls; moderate agreement for myocardial mass; single-operator ground truth with expert review in 150 cases may introduce bias; in external testing: 26% needed minor adjustments, 5% major adjustments, 0.4% crop failures.

Output

CDEs: RDE221, RDE234, RDE220, RDE222

Description: Per-slice segmentation masks for blood pool and myocardium at all phases; derived ventricular volume-time curves and clinical metrics at end diastole and end systole, with mass at end diastole, indexed to BSA.

Recommendation

Suitable for large-scale automated processing of FORCE registry CMR exams to generate standardized ventricular metrics; manual review recommended for cases with poor image quality or flagged for adjustment.

Reproducibility

Public codebase with specified TensorFlow (2.12.0) and Keras (2.8.0) versions; Hyperband used for hyperparameter optimization; deterministic pipeline steps described; average runtime ~26 s per exam on reported setup.

Sustainability

Average processing time per exam 26 s (range 21–32 s); energy consumption not reported.

Use

Intended: Image segmentation

Out-of-scope: Decision support

Excluded: Diagnosis

User

Intended: Other, Researcher

Out-of-scope: Layperson

Excluded: Non-physician provider