Sparse Modeling of Dictionary Learning in Compressive Sensing for Medical Images

Kowsalya A, Hepzibah A. Christinal, D. Abraham Chandy, T. Shekinah, Chandrajit L. BajajWINS 2022

Abstract

This paper investigates enhancing medical imaging efficiency through Compressive Sensing (CS) integrated with deep learning-based dictionary learning. By evaluating reconstruction algorithms on cardiac MRI data, the study identifies Orthogonal Matching Pursuit (OMP) as a superior method, significantly outperforming Least Angle Regression (LAR) in both image quality and computational speed. The results suggest that data-driven sparse modeling can drastically reduce scan times while maintaining diagnostic quality.

Introduction

Magnetic Resonance Imaging (MRI) is a cornerstone of modern diagnostics, yet its reliance on extensive measurements leads to prolonged scan times, high costs, and patient discomfort. Compressive Sensing (CS) addresses this by reconstructing high-quality images from sub-Nyquist sampled data. The efficacy of CS hinges on "sparsity"—the ability to represent complex images using a minimal set of basis functions, or a "dictionary."

Traditional CS methods utilize fixed mathematical transforms like wavelets. However, these often fail to capture the intricate, non-linear features of medical anatomy. This research explores adaptive, data-driven Dictionary Learning (DL) to create more expressive representations for medical image reconstruction.

Methodology

The authors propose a sparse modeling approach centered on Dictionary Learning. Unlike static dictionaries, DL adapts to the specific features of the training dataset.

Conceptual framework of dictionary learning, incorporating various transform types to build an adaptive dictionary.

Key methodological components include:

  • Deep Learning Framework: Utilizing neural architectures to design expressive dictionaries that capture complex image patterns.
  • Data-Driven Tight Frame (DDTF): Employed as a catalyst to enhance the efficiency of the dictionary updating process for large-scale data.
  • Algorithm Comparison: The study evaluates two primary reconstruction techniques: Orthogonal Matching Pursuit (OMP), a greedy iterative algorithm, and Least Angle Regression (LAR), a complex regression-based approach.

Data and Experimental Setup

The study utilized the CAD Cardiac MRI dataset (sourced from Kaggle). To facilitate dictionary training, large-scale MRI images were partitioned into 7,076 smaller image patches. This patch-based strategy allows the model to learn localized features while keeping computational requirements manageable.

Example of a cardiac MRI scan from the CAD dataset used for dictionary training and algorithm evaluation.

Results and Discussion

Performance was measured using Peak Signal-to-Noise Ratio (PSNR) and computation time. The results demonstrated a clear advantage for simpler, faster algorithms in the context of learned dictionaries:

  • OMP (1 atom): The most effective configuration, achieving a PSNR of 33.02 dB in just 0.4 seconds.
  • OMP (2 atoms): Yielded 30.01 dB in 0.8 seconds.
  • LAR (4 atoms): Significantly less efficient, producing a PSNR of 29.12 dB and requiring 5.7 seconds.
  • Thresholding (Baseline): The fastest (0.1s) but lowest quality (25.01 dB), falling below the threshold for reliable medical diagnostics.

One of the most significant insights is the efficiency gap: OMP provided higher accuracy while being approximately 14 times faster than LAR. Furthermore, the success of the 1-atom OMP model suggests that for these specific MRI patches, a highly sparse representation is more robust against noise than more complex multi-atom models.

Conclusion

Sparse modeling using dictionary learning combined with OMP is highly proficient for MRI reconstruction. While the 1-atom OMP model showed the best metrics, the authors note that a balance between sparsity and representation must be maintained to avoid potential bias. Future work will focus on the generalizability of these dictionaries across different imaging modalities (e.g., CT) and anatomical regions to validate their robustness in clinical environments.

Summary updated

BibTeX
@inproceedings{a2022sparse,
  author = {Kowsalya A and Hepzibah A. Christinal and D. Abraham Chandy and T. Shekinah and Chandrajit L. Bajaj},
  title = {{Sparse Modeling of Dictionary Learning in Compressive Sensing for Medical Images}},
  booktitle = {WINS},
  year = {2022},
}
Figure