Evaluation of pseudo-healthy image reconstruction for anomaly detection with deep generative models: Application to brain FDG PET

TL;DR

The study tackles unsupervised anomaly detection in medical imaging by evaluating pseudo-healthy reconstructions when ground-truth lesion masks are unavailable. It introduces a comprehensive simulation-based framework and a 3D VAE trained on healthy brain FDG-PET data to generate pseudo-healthy reconstructions, enabling both reconstruction quality assessment and anomaly localization via atlas regions. The authors demonstrate robust performance on cognitively normal subjects, show generalization to multiple dementia types through simulated anomalies, and validate findings on real ADNI AD patients, supported by latent-space analyses that reveal coherent healthy-data representations and progression trajectories. This framework supports rigorous, automatic validation of pseudo-healthy methods and offers practical steps toward early dementia detection with diffusion model- or diffusion-augmented architectures in the future.

Abstract

Over the past years, pseudo-healthy reconstruction for unsupervised anomaly detection has gained in popularity. This approach has the great advantage of not requiring tedious pixel-wise data annotation and offers possibility to generalize to any kind of anomalies, including that corresponding to rare diseases. By training a deep generative model with only images from healthy subjects, the model will learn to reconstruct pseudo-healthy images. This pseudo-healthy reconstruction is then compared to the input to detect and localize anomalies. The evaluation of such methods often relies on a ground truth lesion mask that is available for test data, which may not exist depending on the application. We propose an evaluation procedure based on the simulation of realistic abnormal images to validate pseudo-healthy reconstruction methods when no ground truth is available. This allows us to extensively test generative models on different kinds of anomalies and measuring their performance using the pair of normal and abnormal images corresponding to the same subject. It can be used as a preliminary automatic step to validate the capacity of a generative model to reconstruct pseudo-healthy images, before a more advanced validation step that would require clinician's expertise. We apply this framework to the reconstruction of 3D brain FDG PET using a convolutional variational autoencoder with the aim to detect as early as possible the neurodegeneration markers that are specific to dementia such as Alzheimer's disease.

Figures (16)

• Figure 1: Hypometabolism simulation pipeline. The intensity of the image from a healthy subject is reduced by a chosen factor in a region associated with a dementia.
• Figure 2: Evaluation framework using simulated images. We simulate an abnormal PET scan $\mathbf{x'}$ from an image of a healthy subject $\mathbf{x}$. If the model works perfectly, the reconstruction $\mathbf{\widehat{x'}}$ should be identical to the original image $\mathbf{x}$.
• Figure 3: Architecture of the 3D convolutional VAE.
• Figure 4: Evolution of the MSE with increasing degrees of hypometabolism simulating AD-like anomalies. We plot the distribution of the MSE between the pseudo-healthy reconstruction and the original image $MSE(\mathbf{x},\mathbf{\widehat{x'}})$ blue, and the MSE between the pseudo-healthy reconstruction and the simulated data $MSE(\mathbf{x'},\mathbf{\widehat{x'}})$ orange. Each MSE is normalized by the average MSE obtained when reconstructing from the original healthy images.
• Figure 5: Example of results obtained from a real image of a CN subject (top row) and an image simulating AD hypometabolism based on the same CN subject (bottom row). For each plane, the first image is the input, the second one the model's reconstruction and the third one the difference (input - reconstruction).