Abstract
Accurate assessment of Acute Myocardial Infarction (AMI) using Cardiac Magnetic Resonance (CMR) imaging requires both precise myocardium segmentation and quantitative $T_2$ relaxation time mapping to identify edema and necrosis. Traditionally, these tasks are performed independently, often leading to $T_2$ maps that are sensitive to noise and lack anatomical consistency. This paper introduces SQNet, a novel dual-task deep learning framework designed to simultaneously perform segmentation and $T_2$ quantification. By employing a hybrid CNN-Transformer architecture and a dedicated Feature Coupling Module (FCM), SQNet leverages the synergy between anatomical structure and tissue characterization. Experimental results demonstrate that SQNet achieves superior segmentation performance (Dice > 89%) and high correlation with gold-standard $T_2$ values ($r=0.93$), while significantly reducing processing time compared to traditional curve-fitting methods.
Introduction
When evaluating patients with AMI, clinicians rely on CMR to delineate the extent of myocardial damage. This involves two primary objectives: segmentation of the heart chambers and myocardium, and the generation of $T_2$ relaxation maps. $T_2$ quantification is particularly challenging because it typically relies on pixel-by-pixel curve fitting (e.g., Nonlinear Least Squares), which is computationally intensive and highly susceptible to image artifacts and noise.
The authors of this study propose that these two tasks are inherently linked. Knowledge of the myocardial boundaries can provide a strong spatial prior for $T_2$ estimation, as pixels within the same tissue region are expected to share similar physical properties. Conversely, the underlying $T_2$ values can help refine segmentation in areas of subtle tissue contrast.
Methods: The SQNet Framework
SQNet is built upon a hybrid architecture that integrates the local feature extraction capabilities of Convolutional Neural Networks (CNNs) with the global context modeling of Transformers.
- Dual-Branch Encoder: The CNN branch captures fine-grained textures and local boundaries, while the Transformer branch uses self-attention to model long-range dependencies across the entire cardiac slice.
- Feature Coupling Module (FCM): This core component fuses features from both branches at multiple stages. It ensures that the segmentation task informs the $T_2$ mapping task by highlighting relevant anatomical regions.
- Task-Specific Decoders: The model utilizes a $T_2$-refine fusion decoder to produce high-quality quantitative maps and a region supervision decoder to ensure sharp, accurate segmentation masks.
- Dynamic Weight Averaging (DWA): To balance the learning process, SQNet employs a dynamic loss function that automatically adjusts the relative importance of the segmentation and quantification losses during training.
Results and Clinical Impact
The model was evaluated on a dataset comprising 120 healthy controls from the CMRxRecon public dataset and 49 in-house AMI cases.
- Segmentation Accuracy: SQNet achieved a Dice similarity coefficient of 89.3% for healthy subjects and 89.2% for AMI patients, outperforming standalone models like Trans-UNet and Swin-UNet.
- Quantification Precision: The $T_2$ values generated by SQNet showed a Pearson correlation of 0.93 with the traditional NLLS "gold standard." Crucially, SQNet was significantly more robust to noise than pixel-wise fitting methods.
- Efficiency: While traditional NLLS fitting takes over 6 minutes per subject, SQNet provides results in seconds, facilitating real-time clinical assessment.
- Radiologist Evaluation: In a blinded study, eight radiologists rated SQNet's outputs as having higher clinical diagnostic quality compared to other state-of-the-art AI methods.
Conclusion
SQNet demonstrates that the simultaneous learning of segmentation and $T_2$ quantification is not only feasible but beneficial. By coupling these tasks, the model achieves greater anatomical consistency and noise robustness in $T_2$ mapping. This integrated approach represents a significant step toward more efficient and reliable automated cardiac MRI analysis in the clinical management of acute myocardial infarction.
Summary updated