Interval Neural Networks as Instability Detectors for Image Reconstructions

Jan MacDonald, Maximilian März, Luis Oala, Wojciech Samek

Introduction

Deep learning models have revolutionized image reconstruction in medical imaging, enabling high-quality results from noisy or incomplete data at unprecedented speeds. However, these models are notoriously fragile. They are susceptible to "instabilities"—tiny, invisible perturbations (adversarial noise) or unfamiliar features (out-of-distribution data) that can lead to "hallucinated" artifacts or the erasure of real pathologies. Because these errors often appear realistic to the human eye, they pose a significant risk in clinical settings. This research proposes that since these instabilities may be an unavoidable trade-off for high performance, we must develop reliable methods to detect them during inference.

Uncertainty Quantification Methods

The authors explore Uncertainty Quantification (UQ) as a safeguard. Instead of providing a single reconstruction, a UQ-enabled model generates a pixel-wise map indicating its confidence level. Three primary methods are compared:

  1. Interval Neural Networks (INN): This approach extends standard networks to output a lower and upper bound for every pixel using interval arithmetic. The uncertainty is defined as the width of this interval.
  2. Monte Carlo Dropout (MCDrop): A common technique where neurons are randomly deactivated during testing. Running the same image multiple times produces a distribution of results; the variance represents the uncertainty.
  3. Mean & Variance Estimation (ProbOut): The network is modified to output both a mean (the reconstruction) and a variance (the uncertainty) for each pixel, trained via a probabilistic loss function.

These methods were evaluated on two tasks: Image Denoising (using the Berkeley Segmentation Dataset) and Limited Angle Computed Tomography (CT) (using the AAPM Low Dose CT Grand Challenge data).

Results and Visual Evidence

The effectiveness of each method was measured by the Pearson correlation between the generated uncertainty map and the actual reconstruction error.

Quantitative Performance

As shown in the table below, INNs demonstrated superior performance in detecting atypical features (Out-of-Distribution) across both tasks. While ProbOut performed slightly better in denoising adversarial detection, INN was the clear leader in the more complex CT reconstruction task.

Quantitative comparison of UQ methods across denoising and CT tasks.

Qualitative Analysis

Visually, the INN uncertainty maps provide a clear "heat map" of potential errors. In the CT experiments, the INN successfully highlighted regions where adversarial noise was added and where out-of-distribution features (such as a silhouette of a dove) were introduced. In contrast, MCDrop often failed to capture these structural instabilities accurately.

Visual comparison of reconstruction and uncertainty maps for CT scans under adversarial and OOD conditions.

Key Insights and Future Work

The study suggests that instability is a fundamental feature of high-performing deep learning models, necessitating a shift toward "self-aware" models that can flag their own failures. INNs proved particularly sensitive to structural changes that other methods missed, providing the spatial precision required for medical diagnostics.

Future research will focus on extending these methods to end-to-end learned reconstructions—mapping raw sensor data directly to images—and addressing the computational overhead associated with training INNs. Ultimately, integrating these uncertainty maps into clinical workflows could provide radiologists with a vital tool for verifying AI-generated reconstructions.

Summary updated

Abstract

This work investigates the detection of instabilities that may occur when utilizing deep learning models for image reconstruction tasks. Although neural networks often empirically outperform traditional reconstruction methods, their usage for sensitive medical applications remains controversial. Indeed, in a recent series of works, it has been demonstrated that deep learning approaches are susceptible to various types of instabilities, caused for instance by adversarial noise or out-ofdistribution features. It is argued that this phenomenon can be observed regardless of the underlying architecture and that there is no easy remedy. Based on this insight, the present work demonstrates, how uncertainty quantification methods can be employed as instability detectors. In particular, it is shown that the recently proposed Interval Neural Networks are highly effective in revealing instabilities of reconstructions. Such an ability is crucial to ensure a safe use of deep learning-based methods for medical image reconstruction.
BibTeX
@inproceedings{macdonald2021interval,
  author = {Jan MacDonald and Maximilian März and Luis Oala and Wojciech Samek},
  title = {{Interval Neural Networks as Instability Detectors for Image Reconstructions}},
  booktitle = {Bildverarbeitung für die Medizin},
  year = {2021},
  doi = {10.1007/978-3-658-33198-6\_79},
}
Figure