Navaneeth Narayanan

Navaneeth Narayanan

Master's Student

Thesis Topic: Deep Convolutional Framelets for Deep Learning-based MRI Reconstruction

Supervisors: Jinho Kim, Prof. Dr. Frederik Laun (UKER, medical supervisor), Prof. Dr. Florian Knoll

Description

Summary: Deep Learning (DL)-based models have significantly improved performance in imaging problems compared to traditional iterative reconstruction methods. Yet, the reasons for the superiority of DL remain partially unclear. To understand the connection between DL and classical signal processing better, Ye et al. extend the concept of “convolutional framelets” and propose a new DL model using multilayer convolution framelets [1]. This model ensures perfect reconstruction (PR) under the ReLU nonlinearity. However, pooling and unpooling layers need high-pass branches to satisfy the PR condition. This insight highlights the limitations of many current deep learning architectures for inverse problems and introduces a new theory for a deep convolutional framelet neural network.

Goal: Implement a DL-based MRI reconstruction model using a Variational Network (VN) [2]. This proposed model will replace max pooling with wavelet pooling in the VN architecture. The performance will be assessed on fastMRI knee and MRCP data, using SSIM or PSNR as quantitative metrics, and compared against the standard VN relative to the ground truth.

Open questions:

  • How does the replacement of max pooling with wavelet pooling influence the overall performance improvement?
  • How does the choice of evaluation data, be it fastMRI knee and MRCP, affect the performance outcomes?

References:

[1] Ye, J. C., Han, Y., & Cha, E. (2018). Deep convolutional framelets: A general deep learning framework for inverse problems. SIAM Journal on Imaging Sciences, 11(2), 991–1048. https://doi.org/10.1137/17M1141771

[2] Sriram, A., Zbontar, J., Murrell, T., Defazio, A., Zitnick, C. L., Yakubova, N., Knoll, F., & Johnson, P. (2020). End-to-End Variational Networks for Accelerated MRI Reconstruction (A. L. Martel, P. Abolmaesumi, D. Stoyanov, D. Mateus, M. A. Zuluaga, S. K. Zhou, D. Racoceanu, & L. Joskowicz, Eds.; Vol. 12262). Springer International Publishing. https://doi.org/10.1007/978-3-030-59713-9