Overview

Schema Version

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

Name

Automated Inline Myocardial Segmentation of Joint T1 and T2 Mapping Using Deep Learning

Link

https://dx.doi.org/10.1148/ryai.220050

Indexing

Keywords: T1 mapping, T2 mapping, cardiac MRI, myocardial segmentation, edge probability estimation, HigherHRNet, mSASHA, inline analysis, AHA 16-segment model

Content: CA, MR

RadLex: RID1399, RID50030

Author(s)

James P. Howard

Kelvin Chow

Liza Chacko

Mariana Fontana

Graham D. Cole

Peter Kellman

Hui Xue

Organization(s)

National Heart and Lung Institute, Imperial College London

National Amyloidosis Centre, Division of Medicine, University College London

Siemens Medical Solutions USA (Cardiovascular MR R&D)

National Heart, Lung, and Blood Institute, National Institutes of Health

Version

1.0

License

Text: MIT License for source code (labeling software and training/assessment)

URL: https://github.com/jphdotam/T1T2_training

Contact

nodnol.stsigoloidrac@hcraeser

Funding

British Heart Foundation (FS/ICRF/22/26039); NIHR Imperial Biomedical Research Centre. Funders had no role in study design, analysis, or drafting.

Ethical review

Ethical approval granted by the UK Health Regulatory Agency (IRAS 243023); waiver of written informed consent due to use of fully de-identified data.

Date

Published: 2022-11-09

Created: 2022-03-09

References

[1] Howard JP, Chow K, Chacko L, Fontana M, Cole GD, Kellman P, Xue H. "Automated Inline Myocardial Segmentation of Joint T1 and T2 Mapping Using Deep Learning". Radiology: Artificial Intelligence. 2023;5(1):e220050. 2023-01-01. doi:10.1148/ryai.220050. PMID: 36721410. PMCID: PMC9885378.

Model

Architecture

Modified HigherHRNet (2D CNN with strided convolutions and multi-resolution aggregation) trained for edge probability estimation of endocardial and epicardial boundaries; five input channels (T1, T2, SR T1w, SR+T2p T2w, PDw) and two output heatmaps.

Availability

Source code for labeling software and for training/testing available on GitHub: https://github.com/jphdotam/T1T2_labeller and https://github.com/jphdotam/T1T2_training

Clinical benefit

Automates myocardial endocardial/epicardial segmentation and provides per-segment T1 and T2 values, enabling rapid, reproducible tissue characterization and reducing manual effort and inter-operator variability.

Clinical workflow phase

Workflow optimization and clinical decision support at the point of image acquisition; inline analysis on MRI scanner console.

Degree of automation

Fully automated inline segmentation and quantification; results displayed automatically on the scanner with overlays for manual QC.

Indications for use

Automated segmentation and segment-wise quantification of myocardial T1 and T2 from joint T1/T2 mapping in adult patients undergoing cardiac MRI; intended for use on MRI scanners in clinical settings.

Input

Coregistered T1 and T2 maps and associated SR T1-weighted, SR+T2p T2-weighted, and proton density–weighted images from mSASHA sequence (1.5 T).

Instructions

Run via Gadgetron InlineAI toolbox (v4.0.1) integrated on the MRI scanner; images reconstructed and maps computed with pixelwise curve fitting; CNN performs boundary detection; per-segment values displayed as bull’s-eye plots and tables on the scanner.

Limitations

All scanners in study were Siemens 1.5-T Aera; generalizability to other vendors/field strengths not established. Only basal/mid/apical short-axis analyzed; long-axis not included. Segment-level analysis may underestimate small subsegment abnormalities. Training/testing labels created with joint viewing of T1 and T2 maps using custom software not typical in routine reporting. Some boundary processing failures occurred (0.8%). Patient-level demographics unavailable due to de-identification.

Output

CDEs: RDE218, RDE1833, RDE223, RDE215

Description: Endocardial and epicardial contours and per-segment (AHA 16-segment) T1 and T2 values, with bull’s-eye plots and tabular report displayed inline on scanner; segmentation overlays for explainability/QC.

Recommendation

Use inline during acquisition to obtain automated per-segment T1/T2 values with contour overlays, enabling rapid review and guiding further acquisitions if needed.

Reproducibility

Agreement with two experts comparable to inter-expert agreement (e.g., T1 native/post-contrast R^2 up to 0.99; T2 R^2 up to 0.92).

Use

Intended: Image segmentation

Out-of-scope: Other

Excluded: Detection and diagnosis

User

Intended: Radiology technologist, Referring physician, Subspecialist diagnostic radiologist

Excluded: Patient, Layperson