Advanced ai–enabled rapid quantitative T2 mapping from needle-free CMR: clinical implications for myocardial edema detection

Background Myocardial edema is a key marker of acute myocardial injury and plays a central role in clinical decision-making. Conventional T2-mapping requires dedicated acquisitions and breath-holds, increasing examination time and motion artifacts. Oxygenation-sensitive (OS-CMR), a needle-free technique, contains intrinsic tissue information that may reflect myocardial relaxation properties [1]. We investigated whether advanced deep learning could generate clinically reliable quantitative T2-maps directly from OS-CMR without additional imaging sequences. Methods Consecutive patients (n = 2,189) who underwent OS-CMR and T2-mapping were included for model development. A residual generative deep learning framework was trained to synthesis quantitative T2-maps directly from OS-CMR images. Physiological fidelity was assessed by comparing T2 decay curves from regions of interest. For clinical validation, eighteen patients with confirmed myocardial edema who underwent were analyzed. Synthetic T2 maps were generated from OS-CMR images in approximately 40 seconds per case and reconstructed in CVI42 for standardized evaluation. Segmental T2 values were extracted using the AHA polar map model (n = 107 segments). The primary endpoint was segment-level agreement between synthetic and reference T2 values, assessed using Pearson correlation, two-way random intraclass correlation (ICC 2,1), and Bland–Altman analysis. The secondary endpoint was patient-level agreement, defined as the mean T2 value across segments. Discriminatory performance for edema detection was evaluated using a predefined clinically established T2 threshold appropriate for the acquisition protocol. Results The novel R-GAN consistently outperformed Pix2Pix (PSNR: 21.56–24.95 vs. 18.80–19.20 dB; SSIM: 0.672–0.739 vs. 0.055–0.659; PCC: 0.862–0.923 vs. 0.800–0.905). Segment-level analysis demonstrated strong quantitative agreement between synthetic and reference T2 values (r = 0.74, ICC = 0.73). Bland–Altman analysis showed minimal bias (+1.1 ms), with limits of agreement (–16.3 to +18.6 ms) within the expected physiological variability of clinical T2 mapping (Figure 1). Mean myocardial T2 values were nearly identical between synthetic and reference maps (51 ± 15 ms vs. 52 ± 9 ms), with preserved spatial distribution of edema on polar and colour maps. For segment-level edema detection, synthetic maps achieved 81.2% sensitivity and 74.1% accuracy. ROC analysis demonstrated strong discriminatory performance (AUC = 0.80) for identifying edema segments (Figure 2), supporting preservation of clinically relevant diagnostic information. Conclusion Advanced deep learning enables rapid (~40 seconds) quantitative T2 mapping from needle-free OS-CMR with strong agreement to conventional T2 and preserved edema detection. This strategy may streamline CMR workflows while maintaining diagnostic integrity. Prospective validation is needed.