Super-Resolution of Brain MRI Images using Overcomplete Dictionaries and Nonlocal Similarity

Yinghua Li, Bin Song, Jie Guo, Xiaojiang Du, Mohsen Guizani

TL;DR

This paper presents a novel single-image super-resolution (SR) method for brain MRIs by combining compressed sensing (CS) with classified overcomplete dictionaries and internal nonlocal self-similarity. By categorizing image patches directly in the measurement domain and leveraging repetitive structures within the image itself, the approach reconstructs high-fidelity, artifact-free medical images without relying on massive, computationally heavy external databases.

What Problem Does This Paper Address?

High-resolution (HR) brain MRI images are critical for accurate medical diagnoses, yet physical, technological, and economic constraints often limit the resolution of raw scans. To bridge this gap, researchers turn to super-resolution algorithms.

Traditional interpolation methods (like bicubic interpolation) are fast but tend to blur high-frequency details and introduce jagged, ringing artifacts. On the other end of the spectrum, machine learning-based approaches can map low-resolution (LR) patches to HR ones, but they require massive external databases, demand high computational power, and risk introducing "hallucinated" or false details not present in the patient's actual anatomy. The authors aim to overcome these limitations by developing a highly accurate, database-light reconstruction method rooted in compressed sensing theory.

Methods

The proposed method splits the super-resolution task into two distinct phases: classified dictionary training and joint image reconstruction.

  1. Classified Dictionary Training: Rather than training a single generic dictionary to represent all image features, the authors classify image blocks into three distinct categories: smooth, texture, and edge blocks. Crucially, this classification happens directly within the measurement domain using covariance matrices, exploiting a proven linear relationship between the cross-covariance matrices of the measurement and frequency domains. Separate sub-dictionaries are then trained for each category using the K-SVD algorithm.
Classified dictionary training pipeline
  1. Nonlocal Self-Similarity Search: Natural images, especially anatomical scans, contain highly repetitive structural features. The algorithm employs a tentative nonlocal search strategy—using a helical square matching pattern for nearby regions and a variable step-size search for distant regions—to locate similar patches within the image itself.
Fig. 3. The similarity of pixels within an image.
  1. Joint CS Reconstruction: During the recovery phase, the target HR image is reconstructed by solving a compressed sensing optimization problem. The objective function enforces two primary constraints: the sparsity of the image patches (via the specialized dictionaries) and their self-similarity with the internally matched blocks. Finally, an Iterative Back Projection (IBP) algorithm refines the global image to ensure it aligns perfectly with the original LR input.
Reconstruction process workflow

Datasets

The performance of the algorithm was validated using three distinct datasets:

  • Brainweb Dataset: A standard benchmark providing synthetic brain MRI images.
  • MRT Dataset: Another widely used synthetic brain MRI dataset.
  • MIDAS Dataset: A real-world dataset consisting of clinical MRI data acquired via a 3T GE scanner, containing scans from 10 normal individuals and 10 schizophrenic patients.

Key Insights and Ideas

  • Measurement Domain Classification: Classifying image patches directly in the measurement domain bypasses the need for full pixel-level reconstruction prior to categorization. This significantly reduces computational overhead while maintaining high classification accuracy.
  • The Power of Internal Priors: Relying purely on external databases can introduce artificial textures. By balancing an external dictionary with internal nonlocal self-similarity, the algorithm acts as a self-regularizing system that respects the true underlying anatomy of the specific scan.

Key Results

The authors compared their method against Bicubic interpolation, SROD, and the deep learning-based BSRCNN across upscaling factors of 2, 3, and 4.

  • Quantitative Improvements: The proposed method consistently outperformed the baseline algorithms in both Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). Across the datasets, PSNR increased by 0.9 to 5.9 dB, and SSIM improved by 0.02 to 0.14. Notably, the performance edge became even more pronounced at higher magnification factors (e.g., 4x upscaling).
  • Qualitative Improvements: Visually, the reconstructed images retained sharper boundaries and richer textures. The typical "staircase" or jagged artifacts seen in interpolation methods were virtually eliminated, providing clearer details for clinical evaluation.

Limitations and Future Work

The primary limitation highlighted in the paper revolves around the measurement domain sampling rate. The authors note that the block classification accuracy is highly sensitive to this rate; if the sampling rate drops below 0.4, the measured value vector shrinks too much, causing a significant drop in classification performance due to information loss.

While explicit future directions are not extensively detailed, adjacent improvements would likely involve optimizing the framework to handle lower sampling rates or extending the joint reconstruction model to handle multi-contrast MRI data simultaneously.

Summary updated

Abstract

Recently, the Magnetic Resonance Imaging (MRI) images have limited and unsatisfactory resolutions due to various constraints such as physical, technological and economic considerations. Super-resolution techniques can obtain high-resolution MRI images. The traditional methods obtained the resolution enhancement of brain MRI by interpolations, affecting the accuracy of the following diagnose process. The requirement for brain image quality is fast increasing. In this paper, we propose an image super-resolution (SR) method based on overcomplete dictionaries and inherent similarity of an image to recover the high-resolution (HR) image from a single low-resolution (LR) image. We explore the nonlocal similarity of the image to tentatively search for similar blocks in the whole image and present a joint reconstruction method based on compressive sensing (CS) and similarity constraints. The sparsity and self-similarity of the image blocks are taken as the constraints. The proposed method is summarized in the following steps. First, a dictionary classification method based on the measurement domain is presented. The image blocks are classified into smooth, texture and edge parts by analyzing their features in the measurement domain. Then, the corresponding dictionaries are trained using the classified image blocks. Equally important, in the reconstruction part, we use the CS reconstruction method to recover the HR brain MRI image, considering both nonlocal similarity and the sparsity of an image as the constraints. This method performs better both visually and quantitatively than some existing methods.
BibTeX
@article{li2019super,
  author = {Yinghua Li and Bin Song and Jie Guo and Xiaojiang Du and Mohsen Guizani},
  title = {{Super-Resolution of Brain MRI Images using Overcomplete Dictionaries and Nonlocal Similarity}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2900125},
}
Figure