Background

Myocardial T1-rho (T1ρ) mapping has shown promise for the identification and quantification of myocardial injury without the need for contrast agent¹. However, its clinical use is hindered by a lack of dedicated analysis tools. Consequently, the reliance on manual processing persists, imposing a dual challenge of radiologist workload and introduction of operator variability. Here, we intend to investigate the feasibility and the benefits of clinically integrated artificial intelligence (AI)-driven analysis for efficient and automated myocardial T1ρ mapping.

Methods

Data: A cohort of 573 patients, composed of healthy cases and patients with various heart conditions, was divided into a train set (N=500, 22% female, age 63 [55-71] years), and a test set (N=73, 68% female, age 55 [41-66] years).
Acquisition: Each patient underwent cardiovascular magnetic resonance on a 1.5T scanner (MAGNETOM Aera, Siemens Healthcare, Erlangen, Germany). Pre-contrast 3-slice T1ρ maps were collected using a single-shot ECG-triggered 2D balanced steady-state free-precession sequence. Five co-registered T1ρ-weighted images were acquired using a T1ρ preparation module² consisting of a 90° tip-down pulse played out along the x-axis to rotate the magnetization, followed by four spin-lock pulses with alternating phases and fixed duration, and two adiabatic 180° pulses. Additionally, a 90° tip-up pulse was then used to return the magnetization to the z-axis. A crusher gradient removed all residual transverse magnetization. For each shot, the spin-lock time was gradually increased (TSL=[0, 10, 20, 35, 50]ms) to generate a T1ρ map through a 2-parameter curve fitting process². In the same subjects, late gadolinium enhancement imaging was subsequently performed 12min post-administration of 0.2mmol/kg gadoteric meglumine using a phase-sensitive inversion recovery sequence³.
Scan parameters: Images were acquired in short-axis during breath-holds with full ventricular coverage. Imaging parameters were: 10-20 slices, 5 signal averages, 1.4x1.4mm2 in-plane resolution, 8mm slice thickness, spin-lock frequency=400Hz, FA=60°, GRAPPA x2, ~160ms acquisition window, TE/TR=1.2/2.8ms, bandwidth=870Hz/pixel, inversion time was determined from a prior scout acquisition to cancel myocardium signal.
Image analysis: The left ventricular (LV) wall was automatically detected from the fourth T1ρ-weighted image (TSL=35ms) using a transformer-based model⁴. A U-Net^5,6 was leveraged to detect right ventricular (RV) insertion points. LV contours and landmarks were propagated onto the T1ρ maps to build a 16-segment American Heart Association (AHA) model⁷, allowing for regional T1ρ value analysis (Fig.1). In both train and test sets, the manual contouring and identification of RV insertion points was performed by expert annotators using CVI42 (Circle Cardiovascular Imaging, Calgary, Canada). The transformer’s hyperparameters were: equally weighted binary cross-entropy, Dice loss, learning rate=10^-3, AdamW optimizer, epochs=77, batch size=8, and weight decay=0.01. A pixel-wise Euclidean distance loss and a 10^-4 learning rate were used for the U-Net.
The proposed method was implemented inline using the Gadgetron framework⁸. Native T1ρ-weighted images and their corresponding T1ρ maps were sent back to the user interface, along with LV wall contours, AHA 16-segment bullseyes, RV insertion points and associated confidence scores.
Experiments: Processing times were measured. Segmentation quality in the test set was assessed using the Dice coefficient. Confidence scores, informing on the probability of accurate detection of landmark insertion points, were retrieved. T1ρ values were quantitatively assessed for both manual and automated processing on patient, slice, and segment levels. Paired Student’s t-tests were used to compare T1ρ values assessed with both processing. T1ρ values difference between slices and between AHA segments were tested using one-way Repeated Measures Analysis of Variance. The concordance between methods was gauged using Bland-Altman analysis and intra-class correlation coefficient (ICC).

Results

Automated analysis of myocardial T1ρ mapping revealed a reduced processing time compared with manual processing (3s vs. 1min51s±22s). Automated segmentation quality yielded favorable results (global Dice 81.2%±6.5). Confidence score for detecting both landmarks reached 94.4% (95% CI: 91.3% - 95.7%) (Fig.2). No significant difference in global T1ρ values was observed between manual (54.9±4.6ms) and automated (55.4±5.1ms) processing (P=0.098). A high level of agreement with expert-derived T1ρ values was found at the patient (ICC=0.85 [0.77-0.90], bias 0.52±5.18ms), slice (ICC=0.80 [0.74-0.84], bias 0.78±8.52ms), and segment (ICC=0.65 [0.62-0.68], bias 0.52±14.48ms) levels (Fig.3). No significant differences in myocardial T1ρ values were found between the automated pipeline and manual processing across the 16 segments defined by the AHA (P>.05) (Fig.4).

Conclusion and Discussion

Automated processing of myocardial T1ρ mapping demonstrates strong agreement with manual processing, bolstered by time efficiency and comparable segmentation quality, highlighting its promising clinical viability for non-invasive, “needle-free”, and rapid on-site assessment of structural heart disease. Further validations in various centers and scanners are warranted.

Acknowledgements

This work was supported by funding from the French National Research Agency under grant agreements Equipex MUSIC ANR-11-EQPX-0030, ANR-21-CE17-0034-01, Programme d’Investissements d’Avenir ANR-10-IAHU04-LIRYC, ANR-22-CPJ2-0009-01, and from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (grant agreement N°101076351).