On instabilities of deep learning in image reconstruction - Does AI come at a cost? (1902.05300v1)

Abstract: 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 unstablemethods for image reconstruction. The instabilities usually occur in several forms: (1) 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 tumour, may not be captured in the reconstructed image and (3) (a counterintuitive type of instability) more samples may yield poorer performance. Our new 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.

• The paper identifies that minor perturbations can induce severe artifacts in deep learning-based image reconstruction.
• The paper reveals that subtle structural changes may remain undetected, raising significant concerns in applications like medical imaging.
• The paper demonstrates that an increased number of samples can unexpectedly worsen performance, deviating from traditional reconstruction methods.

Overview of Deep Learning Instabilities in Image Reconstruction

This paper critically examines the instabilities inherent in employing deep learning methods for image reconstruction in inverse problems. The paper reveals that, despite the significant advancements deep learning has facilitated, its application in image reconstruction introduces various instability issues that merit attention.

Key Instability Phenomena

The research identifies three central forms of instabilities in deep learning-based image reconstruction:

1. Sensitivity to Minor Perturbations: The paper highlights how tiny, almost undetectable perturbations, whether in the image domain or sampling space, can significantly affect the reconstruction quality, often leading to severe artifacts.
2. Lapses in Detecting Structural Changes: Small but crucial structural changes, such as the presence of a tumor, may go unnoticed in reconstructed images, raising concerns over the reliability of deep learning methods in medical imaging where such details are crucial.
3. Counterintuitive Performance with Sample Size: Interestingly, the paper reports scenarios where increasing the number of samples lead to poorer reconstruction performance. This highlights a non-traditional behavior that deviates from classical reconstruction methods where more data typically improves results.

Methodology and Testing

The researchers developed a stability test involving algorithms and software aimed at identifying these instability phenomena. They tested six state-of-the-art networks, each with varying architecture and training data, to observe the extent of instability manifestations. The findings were benchmarked against stable, state-of-the-art methods based on sparse regularization and compressed sensing, ensuring that observed instabilities were attributable to the deep learning techniques and not the inherent problem formulation.

Insights and Observations

The results strongly indicate pervasive instability issues, raising two pivotal questions:

• Does AI inherently bring instability in such high-stake applications?
• Can deep learning models be deemed safe for applications like medical diagnostics, given the risk of undetected artifacts?

The research suggests that, while instabilities do not inherently invalidate deep learning approaches for inverse problems, significant empirical evidence through robust statistical testing is essential for safe deployment, especially in fields such as medical imaging.

Future Directions and Implications

The implications of these findings are profound, advocating for significant research efforts to better understand the causes and mitigate the effects of these instabilities. This includes exploring:

• Further examination of network architectures and training datasets to identify characteristics that reduce instability.
• Development of enhanced algorithms or model designs that inherently account for such perturbations.
• Considerations for retraining models on variable sub-sampling patterns to address performance issues with increased sample sizes.

The paper’s insights into these instability dilemmas underscore an important aspect of the future trajectory for AI in practical applications, urging a rigorous approach to validation and safety assurance in deploying deep learning models. The research emphasizes the necessity for innovative solutions and standards, particularly in high-stakes fields such as healthcare, to ensure AI advancements are safely harnessed.