On instabilities of deep learning in image reconstruction and the potential costs of AI

By Vegard Antun, Francesco Renna, Clarice Poon, Ben Adcock, Anders Christian Hansen. Proceedings of the National Academy of Sciences of the United States of America, 2019

736 citations recorded by the source corpus.

Deep learning, due to its unprecedented success in tasks such as image classification, has emerged as a new tool in image reconstruction with potential to change the field. In this paper, we demonstrate a crucial phenomenon: Deep learning typically yields unstable methods for image reconstruction. The instabilities usually occur in several forms: 1) Certain tiny, almost undetectable perturbations, both in the image and sampling domain, may result in severe artefacts in the reconstruction; 2) a small structural change, for example, a tumor, may not be captured in the reconstructed image; and 3) (a counterintuitive type of instability) more samples may yield poorer performance. Our stability test with algorithms and easy-to-use software detects the instability phenomena. The test is aimed at researchers, to test their networks for instabilities, and for government agencies, such as the Food and Drug Administration (FDA), to secure safe use of deep learning methods.

Subfields: Computational Mechanics, Computer Vision and Pattern Recognition, Radiology, Nuclear Medicine and Imaging

View Paper: https://www.semanticscholar.org/paper/f8a2e5ebf96f55035ebf215156c87a5a9b3be1c5

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computational Mechanics

1. A. Dimakis
2. Giannis Daras
3. Ajil Jalal
4. Stefano Ermon
5. Lei Xing
6. Eric Price
7. Liyue Shen
8. Florian Knoll

In Computer Vision and Pattern Recognition

1. Stefano Ermon
2. Yang Song
3. A. Dimakis
4. Liyue Shen
5. Giannis Daras
6. Lei Xing
7. Ajil Jalal
8. Luc Van Gool

In Radiology, Nuclear Medicine and Imaging

1. Florian Knoll
2. Liyue Shen
3. Lei Xing
4. A. Dimakis
5. Ajil Jalal
6. Giannis Daras
7. John M. Pauly
8. Eric Price

Citations in the corpus

• Robust Compressed Sensing MRI with Deep Generative Priors (2021, background): “However, reconstruction quality is known to suffer when applied out of distribution, and recently has been shown to severely degrade [4, 19] under various types of natural measurement and anatomy perturbations.”
• Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction (2021, methodology): “Neural network models can be unstable as demonstrated via adversarial perturbation studies [64].”
• NeRP: Implicit Neural Representation Learning with Prior Embedding for Sparsely Sampled Image Reconstruction (2021, background): “…can be a bottleneck, the reconstructions may not be robust when deployed to unseen subjects, the reconstructions can be unstable with subtle yet significant structural changes such as tumor growth, and there can also be difficulties generalizing to different image modalities or anatomical sites [16].”
• Solving Inverse Problems in Medical Imaging with Score-Based Generative Models (2021, background): “This prevents models from generalizing effectively to new measurement processes, leading to counter-intuitive instabilities such as more measurements causing worse performance (Antun et al., 2020).”
• FISTA-Net: Learning a Fast Iterative Shrinkage Thresholding Network for Inverse Problems in Imaging (2020, background): “capturing critical features of data, its stability is usually unsatisfactory [27].”
• Democratising deep learning for microscopy with ZeroCostDL4Mic (2021, methodology): “However, using pretrained models bears the risk of predicting incorrect structures and artefacts if they are used on unseen data too dissimilar to the data they were originally trained on 24,25 .”