Hallucination Index: An Image Quality Metric for Generative Reconstruction Models

Matthew Tivnan, Siyeop Yoon, Zhennong Chen, Xiang Li, Dufan Wu, Quanzheng Li

Abstract

Generative image reconstruction algorithms, such as measurement-conditioned diffusion models, are increasingly popular in the field of medical imaging. These powerful models can transform low signal-to-noise ratio (SNR) inputs into outputs with the appearance of high SNR. However, the outputs can have a new type of error called hallucinations. In medical imaging, these hallucinations may not be obvious to a radiologist but could cause diagnostic errors. Generally, hallucination refers to error in estimation of object structure caused by a machine learning model, but there is no widely accepted method to evaluate hallucination magnitude. In this work, we propose a new image quality metric called the hallucination index. Our approach is to compute the Hellinger distance from the distribution of reconstructed images to a zero-hallucination reference distribution. To evaluate our approach, we conducted a numerical experiment with electron microscopy images, simulated noisy measurements, and applied diffusion-based reconstructions. We sampled the measurements and the generative reconstructions repeatedly to compute the sample mean and covariance. For the zero-hallucination reference, we used the forward diffusion process applied to ground truth. Our results show that higher measurement SNR leads to lower hallucination index for the same apparent image quality. We also evaluated the impact of early stopping in the reverse diffusion process and found that more modest denoising strengths can reduce hallucination. We believe this metric could be useful for evaluation of generative image reconstructions or as a warning label to inform radiologists about the degree of hallucinations in medical images.

Introduction

In medical imaging, clinicians often face a trade-off between image quality and acquisition constraints, such as radiation dose or scan time. This frequently results in low signal-to-noise ratio (SNR) data. Modern generative models, including GANs and diffusion models, have shown remarkable ability to "clean" this data, producing sharp, high-contrast images that mimic high-SNR acquisitions.

However, these models do not merely denoise; they often "fill in the blanks" by generating plausible but fake structural details—a phenomenon known as hallucination. In a diagnostic setting, a hallucinated lesion or vessel could lead to catastrophic clinical errors. Despite the risk, the field lacks a standardized quantitative metric to distinguish between real structural recovery and AI-invented artifacts. This paper introduces the Hallucination Index (HI) to address this gap.

Methodology

The proposed Hallucination Index (HI) is rooted in information theory and probability distribution comparison. The core idea is to define hallucinations by negation: identifying what structural information is added by the model that was not present in the original ground truth, even when accounting for noise.

The Conceptual Framework

We define the reconstruction process as a mapping from measurements $y$ to a reconstructed image $\hat{x}$. The model's output is characterized by a distribution $p_\theta(\hat{x}|x)$. To isolate hallucinations, we compare this to a Zero-Hallucination Reference distribution $q(\hat{x}|x)$. This reference is created by taking the ground truth $x$ and applying the same degradation (noise/blur) present in the measurements, but without the generative model's "creative" reconstruction step. In diffusion models, this corresponds to the forward diffusion process.

Conceptual framework and mathematical definition of the Hallucination Index (HI).

The Metric

The difference between these two distributions is quantified using the Hellinger distance. The index $H$ is calculated as:

$$H = \sqrt{1 - \int \sqrt{p_\theta(\hat{\mathbf{x}}|\mathbf{x})}\sqrt{q(\hat{\mathbf{x}}|\mathbf{x})} d\hat{\mathbf{x}}}$$

A score of 0% indicates the AI output perfectly matches the noisy reference (no hallucination), while 100% indicates the distributions are entirely disjoint, implying the AI has completely invented the image structure. To make this calculation tractable for high-dimensional images, the authors approximate the distributions as multivariate Gaussians, allowing for a closed-form solution based on sample mean and covariance.

Experimental Setup

The metric was validated using the Machine Intelligence from Cortical Networks (MICrONS) dataset, which contains high-resolution electron microscopy (EM) images of cortical neurons. This dataset is particularly challenging because neural connectivity (the connectome) is highly sensitive to small structural errors; even minor hallucinations can lead to incorrect mapping of neural circuits. The study used 10,000 images for training and 1,000 for evaluation, all resized to 256x256 pixels.

Key Results

The experiments revealed several critical insights into the behavior of generative reconstruction models:

  1. SNR Dependency: As the input measurement SNR increases, the Hallucination Index decreases. With more "real" data available, the model relies less on its learned priors to guess missing structures.
  2. The MSE/HI Trade-off: Standard metrics like Mean Squared Error (MSE) can be deceptive. While running more iterations of a diffusion model typically reduces MSE and makes images look "cleaner," it simultaneously increases the Hallucination Index. This suggests that the perceived "sharpness" is often achieved through hallucination.
  3. Early Stopping: The authors found that "early stopping" in the reverse diffusion process—halting before the image is fully denoised—significantly lowers the HI. This keeps the image slightly noisier but ensures it remains more "honest" to the original measurements.
Visual comparison of ground truth, simulated noisy measurements, and the generative reconstruction process.

Discussion and Conclusion

While the Hallucination Index provides a powerful new tool for evaluation, it has limitations. The current implementation assumes Gaussian, shift-invariant distributions, which may not capture complex non-linear errors. Furthermore, for models designed to produce perfectly noiseless images, the index may default toward 100% because a noiseless image and a noisy reference have zero overlap, necessitating further refinement in how the reference is defined.

Ultimately, the Hallucination Index serves as a vital "warning label" for medical imaging. By providing a quantitative measure of AI-induced artifacts, it allows researchers to optimize models for fidelity rather than just aesthetics and helps clinicians understand the level of trust they should place in an AI-enhanced image.

Summary updated

BibTeX
@inproceedings{tivnan2024hallucination,
  author = {Matthew Tivnan and Siyeop Yoon and Zhennong Chen and Xiang Li and Dufan Wu and Quanzheng Li},
  title = {{Hallucination Index: An Image Quality Metric for Generative Reconstruction Models}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2024},
  url = {https://openreview.net/forum?id=6ZttEKZcuG},
}
Figure