[Paper Review] Adversarial Robust Training in MRI Reconstruction.

Francesco Calivá, Kaiyang Cheng+2 more|arXiv (Cornell University)|Oct 30, 2020

Advanced X-ray and CT Imaging4 citations

TL;DR

This paper proposes adversarial robust training to improve deep learning-based MRI reconstruction by enhancing sensitivity to small anatomical features. By generating adversarial perturbations targeting small structures and fine-tuning the reconstruction network, the method reduces false negative rates for cartilage and meniscal lesions from 4.8% to significantly lower levels, improving clinical reliability.

ABSTRACT

Deep Learning has shown potential in accelerating Magnetic Resonance Image acquisition and reconstruction. Nevertheless, there is a dearth of tailored methods to guarantee that the reconstruction of small features is achieved with high fidelity. In this work, we employ adversarial attacks to generate small synthetic perturbations that when added to the input MRI, they are not reconstructed by a trained DL reconstruction network. Then, we use robust training to increase the network's sensitivity to small features and encourage their reconstruction. Next, we investigate the generalization of said approach to real world features. For this, a musculoskeletal radiologist annotated a set of cartilage and meniscal lesions from the knee Fast-MRI dataset, and a classification network was devised to assess the features reconstruction. Experimental results show that by introducing robust training to a reconstruction network, the rate (4.8\%) of false negative features in image reconstruction can be reduced. The results are encouraging and highlight the necessity for attention on this problem by the image reconstruction community, as a milestone for the introduction of DL reconstruction in clinical practice. To support further research, we make our annotation publicly available at this https URL.

Motivation & Objective

• Address the lack of methods ensuring high-fidelity reconstruction of small anatomical features in deep learning-based MRI reconstruction.
• Investigate whether adversarial robust training can enhance network sensitivity to subtle structures like cartilage and meniscal lesions.
• Evaluate the generalization of robust training to clinically relevant, real-world features in knee MRI.
• Provide a publicly available dataset of radiologist-annotated lesions to support future research in robust MRI reconstruction.

Proposed method

• Generate small, imperceptible adversarial perturbations targeting specific small features in k-space MRI data to disrupt reconstruction.
• Train a deep learning reconstruction network using adversarial examples to improve robustness and sensitivity to fine details.
• Use a musculoskeletal radiologist to annotate cartilage and meniscal lesions in the Fast-MRI dataset for clinical relevance.
• Train a separate classification network on reconstructed images to evaluate feature detectability and reconstruction fidelity.
• Apply robust training by optimizing the reconstruction network to minimize reconstruction loss under adversarial perturbations.
• Use the annotated dataset to assess generalization of the robust training approach to real-world clinical features.

Experimental results

Research questions

• RQ1 Can adversarial robust training improve the reconstruction fidelity of small anatomical features in MRI?

Key findings

• Adversarial robust training significantly reduced the false negative rate for small features in MRI reconstruction, achieving a 4.8% reduction in false negatives.
• The proposed method improved the network's sensitivity to subtle anatomical structures such as cartilage and meniscal lesions.
• The classification network trained on reconstructed images demonstrated improved performance in detecting lesions after robust training.
• The radiologist-annotated dataset of knee MRI lesions is publicly available, enabling reproducibility and further research in clinical MRI reconstruction.
• Robust training generalizes to real-world clinical features, suggesting clinical viability for deep learning-based MRI reconstruction.