논문 출처 ➡️ https://www.mdpi.com/2076-3417/12/22/11758
A Deep Learning Approach to Upscaling “Low-Quality” MR Images: An In Silico Comparison Study Based on the UNet Framework
MR scans of low-gamma X-nuclei, low-concentration metabolites, or standard imaging at very low field entail a challenging tradeoff between resolution, signal-to-noise, and acquisition duration. Deep learning (DL) techniques, such as UNets, can potentially
www.mdpi.com
Input Data and Preprocessing
The study used both low-quality (LQ) and high-quality (HQ) magnetic resonance imaging (MRI) data. Since matched LQ-HQ datasets were not publicly available, the authors generated synthetic datasets using 436 images from the OASIS project. These images were augmented to create 2616 datasets. Each dataset included synthetic high-quality images (synHQ), low-quality images (synLQ), and a high-quality complementary prior (HQCP) used in certain scenarios.
The LQ images were created by adding Gaussian noise and downsampling the original HQ images using k-space truncation at different resolutions (1/8, 1/4, 1/2). The synLQ images were interpolated back to the original matrix size (176x208) using nearest-neighbor interpolation before feeding them into the models.
Data Processing Scenarios
The study tested two main acquisition scenarios:
- With-Priors (HQCP): Both LQ images and a high-quality complementary prior (HQCP) image (same subject, higher resolution or contrast) were used as inputs to upscale LQ images.
- Without-Priors (NoCP): Only LQ images were used as input, without any complementary HQ images.
Models Used
The authors used three different UNet variants:
- Dense UNet (DUNet): Uses dense connections between layers to carry outputs from one layer to all subsequent layers, enabling learning of more complex features.
- Robust UNet (RUNet): Uses residual blocks and batch normalization to improve convergence and feature extraction, using residual additions to enhance learning.
- Anisotropic UNet (AUNet): This architecture is based on bottleneck and residual cores, designed to handle non-isotropic images, though only 2D operations were used for isotropic images in this study.
Methods
The authors trained the networks on synthetic datasets, using HQ images as ground truth. Each network was trained for 100 and 1000 epochs using the Adam optimizer and a learning rate of 0.001. Mean squared error (MSE) was used as the loss function. The models were trained and evaluated on two scenarios (With-Priors and Without-Priors) and three different downsampled LQ image sizes (1/8, 1/4, and 1/2 of the full matrix size). The key metrics for evaluating the model's performance were MSE, Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Mean Intensity Error (MIE).
Results and Conclusions
- Performance Across UNets: The study found that the differences in performance between DUNet, RUNet, and AUNet were not statistically significant, indicating that the architectural details of these UNets may not play a critical role in the upscaling task. All three networks showed similar convergence and performance in upscaling LQ images.
- With-Prior vs. Without-Prior: As expected, the With-Priors scenario (where HQCP images were available) led to significantly better results than the Without-Priors scenario. The complementary HQCP images helped the UNet recover finer details lost in LQ images.
- Importance of Acquisition Matrix Size: Larger matrix sizes for LQ images (e.g., 1/2 downsampled) resulted in better upscaling performance compared to smaller ones (e.g., 1/8 downsampled). In the Without-Priors scenario, lower-resolution inputs led to worse reconstruction of fine details, such as small lesions, which were not recoverable without complementary HQ data.
- Statistical Analysis: Mixed-effects modeling showed that the type of UNet, matrix size, and their interaction were statistically significant factors in performance. However, all networks performed similarly when sufficient training data were provided.
Summary
The study supports the use of deep learning, specifically UNet-based architectures, for upscaling low-quality MRI images. The availability of high-quality complementary images (With-Priors) significantly improved the results. The research also highlights that network architecture might not be the most critical factor, as long as the model is trained with adequate data and proper training protocols. This study advances the potential integration of DL-based methods in improving low-quality MRI acquisitions, particularly when dealing with low-resolution data.
