Abstract
In clinical MRI, a trade-off often exists between scan time and spatial resolution, frequently resulting in "thick-slice" images that lack detail in non-primary planes. This paper introduces TSCNet (Two-stage Self-supervised Cycle-consistency Network), a framework designed to reconstruct high-resolution, thin-slice MR images from low-resolution data. By employing a two-stage self-supervised strategy—leveraging cross-plane synthesis and axial cycle-consistency—TSCNet achieves performance comparable to fully supervised models without requiring paired ground-truth training data.
The Challenge of Thick-Slice MRI
Radiologists often acquire thick-slice MRI scans to maintain a high signal-to-noise ratio and keep acquisition times manageable. While these scans provide high resolution in the primary axial plane, they appear blurry in sagittal and coronal views. This anisotropy limits the utility of the images for detailed diagnosis and automated analysis. While deep learning offers a solution through slice interpolation, most models require large datasets of paired "thick" and "thin" slices, which are rare in clinical practice. Existing self-supervised methods often fail to maintain anatomical accuracy because they ignore the 3D context of the human body.
Proposed Method: TSCNet
TSCNet addresses these limitations through a novel Two-stage Self-supervised Learning (SSL) approach that utilizes the inherent properties of MRI data.
Stage 1: Cross-Plane Synthesis
The first stage exploits the fact that while axial resolution may be low, the sagittal and coronal planes in the same scan are often densely sampled. The authors downsample these dense planes to create synthetic "thick-slice" pairs. A Residual Dense Network (RDN) is then trained on these pairs to learn basic human anatomy and the fundamentals of slice interpolation.
Stage 2: Cycle-Consistency Refinement
To adapt the model specifically to the axial direction, the second stage introduces a cycle-consistency constraint using triplets of real axial slices ($S_i, S_{i+1}, S_{i+2}$):
- The network interpolates intermediate slices $S_{i+0.5}$ and $S_{i+1.5}$.
- It then attempts to reconstruct the original central slice $S_{i+1}$ from these two new interpolated slices.
- A reconstruction loss (MSE) and an adversarial loss (via a discriminator) are used to ensure the results are both mathematically accurate and visually realistic, avoiding the common pitfall of over-smoothing.
Experimental Setup
The model was evaluated using 64 T1-weighted brain MR images from the ADNI dataset. Ground truth images were isotropic ($1 \times 1 \times 1 \text{ mm}^3$), and input data was simulated by downsampling the axial direction by a factor of 2.
Key Results and Discussion
TSCNet demonstrated superior performance compared to traditional interpolation and existing self-supervised methods, nearly matching the performance of fully supervised benchmarks.
Quantitative Performance
As shown in the table below, TSCNet significantly outperforms standard self-supervised approaches like EDSSR in both Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM).
| Algorithm | Category | PSNR (dB) | SSIM |
|---|---|---|---|
| Trilinear | Conventional | 34.03 | 0.9139 |
| EDSSR | SSL | 35.44 | 0.9344 |
| TSCNet (Ours) | SSL | 36.14 | 0.9419 |
| FSIN | Fully Supervised | 36.53 | 0.9447 |
Qualitative Analysis
Visual inspection confirms that TSCNet effectively reconstructs sharp tissue boundaries and realistic anatomical details. By leveraging the "cycle" as a free supervisor, the model learns the specific characteristics of axial slices without ever seeing a high-resolution axial scan during training.
Conclusion
TSCNet provides a robust framework for high-resolution MRI reconstruction in data-scarce environments. By cleverly utilizing cross-plane anisotropy and cycle-consistency, it bridges the gap between self-supervised and fully supervised learning. Future work may explore the generalization of this method to other anatomical regions and the integration of Transformer-based architectures to further enhance reconstruction quality.
Summary updated