Background/Objectives: The composition of atherosclerotic plaques is increasingly recognized as a key factor determining cardiovascular risk. Features such as intraplaque hemorrhage, a necrotic lipid core, and the integrity of the fibrous cap are strongly associated with plaque instability and the occurrence of adverse clinical events. Magnetic resonance imaging (MRI) allows for non-invasive characterization of plaque microstructure through quantitative mapping of T1 and T2 relaxation times; however, image noise may limit the accuracy of these measurements. Methods: In this experimental study, a total of 15 ex vivo atherosclerotic plaque samples were imaged using a 1.5T scanner with a fast spin-echo sequence featuring variable repetition times (TR: 200–12,000 ms) and echo times (TE: 21–240 ms) to obtain T1 and T2 maps. An Attention–Residual–Dense U-Net neural network was trained on pairs of noisy and reference images to reduce Rician noise while preserving structural details. Results: The 15 samples examined exhibited T1 values ranging from 1768 to 3294 ms and T2 values ranging from 138 to 202 ms, which were shorter than those for water (T1: 3323 ms; T2: 114 ms), which is consistent with the presence of collagen, lipids, and mineral deposits. Variability among samples reflected differences in composition, with the shortest relaxation times suggesting advanced calcifications. The application of deep learning methods allowed for a threefold improvement in the signal-to-noise ratio (SNR) while preserving the microarchitecture of the lamina. Conclusions: Quantitative T1/T2 mapping combined with deep learning-based image enhancement methods constitutes a robust tool for high-resolution characterization of atherosclerotic plaque composition under ex vivo conditions. The results obtained indicate the potential for translating this method to in vivo studies to better detect tissue heterogeneity and features associated with plaque instability.
Myśliwiec et al. (Sun,) studied this question.