QN-Mixer: A Quasi-Newton MLP-Mixer Model for Sparse-View CT Reconstruction

Ishak Ayad, Nicolas Larue, Maï K. NguyenCVPR 2024 · DOI · arXiv

Abstract

Sparse-view Computed Tomography (CT) is a critical technique for reducing patient radiation exposure, but it introduces significant challenges due to the ill-posed nature of the reconstruction problem. This paper introduces QN-Mixer, a novel deep unrolling framework that combines second-order optimization with a hybrid neural architecture. By utilizing a latent Quasi-Newton (BFGS) update and an "Incept-Mixer" regularizer, QN-Mixer achieves high-quality medical reconstructions with superior computational efficiency and fewer iterations than current state-of-the-art methods.

Introduction

Computed Tomography (CT) is indispensable for medical diagnosis, yet the ionizing radiation from X-rays poses health risks. Adhering to the ALARA (As Low As Reasonably Achievable) principle, clinicians often employ sparse-view CT, which utilizes fewer X-ray projections. However, this reduction in data leads to severe "streaking" artifacts in standard reconstructions.

Existing deep learning approaches typically follow two paradigms: post-processing (denoising low-quality images) or deep unrolling (mapping iterative optimization steps into neural layers). While deep unrolling is generally more effective, standard first-order methods (like gradient descent) converge slowly and often rely on CNNs that lack long-range dependency modeling or memory-intensive Transformers. QN-Mixer addresses these limitations by introducing second-order optimization into the unrolling framework.

Methods

QN-Mixer employs a two-pronged strategy to enhance reconstruction efficiency and quality:

Latent Quasi-Newton (BFGS) Unrolling

Unlike first-order methods that rely on simple gradient descent, QN-Mixer utilizes the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm. As a second-order method, BFGS approximates the Hessian matrix to determine more optimal update steps. To manage the high memory cost associated with Hessian matrices in high-resolution imaging, the authors project gradients into a latent space. The BFGS update is performed within this compressed space before being decoded back to the image domain, providing superlinear convergence without quadratic memory growth.

The QN-Mixer unrolling framework showing the iterative BFGS update process.

Incept-Mixer Regularization

The model incorporates a custom neural block termed Incept-Mixer as its regularization term. This hybrid architecture combines Inception modules—which excel at capturing local textures and multi-scale features—with MLP-Mixer layers that use an attention-like mechanism to capture long-range global dependencies. This ensures that both minute lesions and large-scale anatomical structures are accurately preserved.

Key Results

The model was rigorously evaluated on the AAPM (2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge) and DeepLesion datasets.

  • Quantitative Performance: On the AAPM dataset (32 views), QN-Mixer achieved a PSNR of 39.51 dB, significantly outperforming the previous state-of-the-art RegFormer (38.71 dB) and the standard LEARN model (37.58 dB).
  • Efficiency: QN-Mixer requires only 14 iterations to reach peak performance, compared to the 30+ iterations often needed by first-order models. It also demonstrated a 1.3x reduction in memory usage compared to transformer-based models like RegFormer.
  • Robustness: In Out-of-Distribution (OOD) testing involving unexpected artifacts, QN-Mixer maintained high structural similarity (SSIM) where first-order networks showed significant degradation.
Comparison of QN-Mixer against post-processing and first-order unrolling methods.

Discussion and Conclusion

QN-Mixer represents a significant advancement in sparse-view CT reconstruction by successfully integrating second-order optimization into a deep unrolling framework. While the latent BFGS approach is highly efficient, future work will explore scaling to 3D CT volumes and developing unsupervised variants to eliminate the need for paired full-dose training data. Currently, the model's ability to balance local detail with global structure makes it a promising candidate for clinical low-dose CT applications.

Summary updated

BibTeX
@inproceedings{ayad2024qn,
  author = {Ishak Ayad and Nicolas Larue and Maï K. Nguyen},
  title = {{QN-Mixer: A Quasi-Newton MLP-Mixer Model for Sparse-View CT Reconstruction}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2024},
  doi = {10.1109/cvpr52733.2024.02392},
}
Figure