Rapid Fire Session

Session: Rapid Fire Session 5: Quantitative MRI and Artificial Intelligence (Monitor 8)

Development and Validation of Deep-Learning Based T1 Mapping Quantification and Risk Prediction in Cardiac Amyloidosis

Saturday, February 7, 2026

10:50 AM - 10:57 AM

Location: Galapagos / Aruba / Ilhabela (2nd Lower Level)

Presenting Author(s)

Anna Y. Birns

Computer Science Co-op Student
Libin Cardiovascular Institute, University of Calgary
Calgary , Alberta, Canada

Primary Author(s)

Anna Y. Birns

Computer Science Co-op Student
Libin Cardiovascular Institute, University of Calgary
Calgary , Alberta, Canada

Co-Author(s)

CJ

Cory R. Jensen, BSc

Cardiac Machine Learning Developer
Libin Cardiovascular Institute, University of Calgary, Canada
Justin J. Tse, PhD, MSc, BSc

Research Associate
University of Calgary
Calgary, Alberta, Canada
GW

Geneva C. White

Summer Student
Libin Cardiovascular Institute, University of Calgary, Canada
FH

Fereshteh Hasanzadeh, MSc

Masters Student
University of Calgary
calgary, Alberta, Canada
JF

Jacqueline Flewitt, MSc

Manager of Strategic Partnerships
Libin Cardiovascular Institute; University of Calgary
Calgary, Alberta, Canada
TB

Talia A. Beckie, BEng

Cardiac Imaging Systems Engineer
Libin cardiovascular Institute, University of Calgary
Calgary, Alberta, Canada
SD

Steven Dykstra, MSc, BEng

PhD Candidate
University of Calgary
Calgary, Alberta, Canada
SR

Sandra Rivest, RN

Research Coordinator
Libin Cardiovascular Institute; University of Calgary
Calgary, Alberta, Canada
YF

Yuanchao Feng, PhD

Senior AHS analyst
Libin Cardiovascular Institute; University of Calgary
Calgary, Alberta, Canada
JN

Julian B. Neves, MSc

Junior Data Engineer
Libin Cardiovascular Institute, University of Calgary, Canada
AH

Andrew G. Howarth, MD, PhD

Associate Professor
Libin Cardiovascular Institute; University of Calgary
Calgary, Alberta, Canada
CL

Carmen P. Lydell, MD

Radiologist
Stephenson Cardiac Imaging Centre, Libin Cardiovascular Institute of Alberta, University of Calgary
Calgary, Alberta, Canada
LK

Louis Kolman, MD

Clinical Assistant Professor
Libin Cardiovascular Institute; University of Calgary
Calgary, Alberta, Canada
RM

Robert JH Miller, MD

Cardiologist
Libin Cardiovascular Institute, University of Calgary, Canada
NF

Nowell M M. Fine, MD

Associate Professor
Libin Cardiovascular Institute of Alberta, University of Calgary
Calgary, Alberta, Canada
Dina Labib, MD, PhD, FSCMR

Associate Scientific Director, Personalized Diagnostics Program; Adjunct Assistant Professor
University of Calgary
Calgary, Alberta, Canada
James A. White, MD

Professor
Libin Cardiovascular Institute; University of Calgary
Calgary, Alberta, Canada
AA

Augustine Amakiri, PhD

Data Analyst
University of Calgary, Canada
MK

Melanie King, PhD

Director, Program Development and Implementation
Libin Cardiovascular Institute, University of Calgary, Canada
MB

Michael S. Bristow, MD

Clinical Assistant Professor
Libin Cardiovascular Institute of Alberta, University of Calgary, Alberta, Canada
MG

Marco J.W Götte, MD, PhD

MD, PhD
Stephenson Cardiac Imaging Centre
Calgary, Alberta, Canada

Background: T1 mapping has established prognostic value in multiple cardiac diseases, including amyloidosis. Manual quantification is time-consuming and poorly reproducible. Several deep learning segmentation algorithms have been developed, trained mostly from healthy hearts, performing poorly when applied to advanced disease. In this study, we aimed to develop and validate a deep learning model for T1 map segmentation in diseased hearts, and study prognostic value of derived regional metrics in cardiac amyloidosis.

Methods: Using a 2D U-Net architecture and a 70 training/15 validation/15 testing split, we developed a segmentation algorithm for endocardial and epicardial contouring for MOLLI T1 mapping (native, post-contrast, and synthetic ECV). Manually annotated contouring from multi-slice short axis maps was performed in 350 subjects, inclusive of healthy volunteers and four cardiomyopathy states: hypertrophic cardiomyopathy, Fabry disease, cardiac amyloid, and ischemic cardiomyopathy. Overall model performance was evaluated via DICE scores on the hold-out test set. Additionally, a separate U-Net model was trained for the prediction of right ventricular insertion (RVI) points to automate division of the myocardium into AHA segments. Automated segmentations were then used to derive unique T1 metrics in a cohort with confirmed cardiac amyloidosis (n=135). Conventional global mean values as well as i) basal to apical slice mean gradient, and ii) endocardial to epicardial zone mean gradients were computed from the basal, mid and apical native T1 and ECV maps. Univariable and multivariable Cox models were performed to test associations between imaging markers and a composite outcome of all-cause death and heart failure admission.

Results: The deep learning model achieved high segmentation accuracy, with DICE scores ranging from >90% for endocardium and 82-88% for the epicardium across cohorts. Mean pixel distance difference for ground truth and predicted RVI was 2.9 mm. In cardiac amyloidosis, univariable analyses revealed multiple native T1 and ECV metrics to be significantly associated with adverse outcomes (Table 1). Adjusting for age, sex, and atrial fibrillation, automated mean global ECV was independently associated with worse survival (Figure 1; HR [95% CI] 1.051 [1.022-1.080]; p< 0.001), while mean native T1 lost significance. However, reduced native T1 base/apex gradient, denoting more distal infiltration of the LV myocardium, was associated with worse survival (HR 1.063 [95% CI: 1.012–1.117]; p=0.015). Representative examples from outcome-positive and outcome-negative patients are shown in Figure 2.

Conclusion: We developed a fully automated, deep learning-based algorithm for native T1 map and ECV quantification maintaining excellent performance across healthy and diseased hearts. Generated conventional and novel, gradient-based metrics show strong potential to improve risk modelling in patients with cardiac amyloidosis.

Univariable survival associations of T1 Metrics derived from automated T1 mapping segmentations and the composite outcome in patients with cardiac amyloidosis

Characteristic	HR (95% CI)	p-value
Mean global native T1, per 1 ms	1.002 (0.999-1.005)	0.21
Mean endocardial native T1, per 1 ms	1.002 (0.999-1.006)	0.13
Mean epicardial native T1, per 1 ms	1.003 (1.000-1.006)	0.092
Mean basal native T1, per 1 ms	1.002 (0.999-1.006)	0.25
Mean apical native T1, per 1 ms	1.004 (1.001-1.008)	0.006
Native T1 endocardial/epicardial gradient, per 1%	0.953 (0.785-1.156)	0.63
Native T1 base/apex gradient, per 1 decrease%	1.059 (1.021–1.100)	0.002
Mean global ECV T1, per 1 ms	1.040 (1.015-1.065)	0.002
Mean endocardial ECV, per 1 %	1.037 (1.012-1.061)	0.003
Mean epicardial ECV, per 1 %	1.041 (1.016-1.067)	0.001
Mean basal ECV, per 1 %	1.038 (1.014-1.062)	0.002
Mean apical ECV, %	1.033 (1.011-1.056)	0.003
ECV endocardial/epicardial gradient, per 1%	1.015 (0.954-1.081)	0.64
ECV base/apex gradient, per 1%	1.002 (0.987-1.018)	0.77

CI indicates confidence interval; HR, hazard ratio; ECV, extracellular volume.
Figure 1. Forest plot of a multivariable Cox model for associations with the composite outcome
Figure 2. Representative example of an outcome-negative and outcome-positive patient with predicted segmentations of native T1 and synthetic ECV maps, as well as derived T1 metrics