Ninon Burgos

Deep generative models for the detection of anomalies in brain PET images

Neuroimaging offers an unmatched description of the brain’s structure and physiology, which explains its crucial role in the understanding, diagnosis, and treatment of neurological disorders. However, identifying subtle pathological changes simply by looking at images of the brain can be a difficult task, and in clinical practice, many imaging modalities such as positron emission tomography (PET) are analysed visually, with sensitivity and specificity greatly depending on the observer’s experience. Images are often transported to a standard space where they can be visually or quantitatively compared to images of normal controls, but this approach suffers from a lack of sensitivity due to variability between subjects in non-pathological areas.

During my postdoc, I proposed an approach to mitigate this limitation that consists of generating a healthy-looking image specific to the patient under investigation, based on a registration and fusion algorithm: when compared to the real image of the patient, this pseudo-healthy model can be used to detect the areas of the image that show abnormalities (more details on this page). This line of work has since progressed from this traditional image processing technique to deep generative models trained exclusively on healthy subjects, including variational autoencoders and more recent architectures derived from diffusion models, where the patient’s image is compared at inference to its pseudo-healthy reconstruction to highlight areas presenting anomalies (see figure below).

Anomaly detection framework using variational autoencoders
Anomaly detection framework using variational autoencoders. The network, trained only on healthy images, reconstructs a pseudo-healthy version of a patient's image; comparing the two highlights potential anomalies.

A first key contribution, constituting the core of Ravi Hassanaly’s PhD thesis (2020–2024), was twofold. First, we introduced a simulation-based evaluation framework to assess pseudo-healthy reconstruction methods in the absence of ground truth anomaly masks (Hassanaly et al., MELBA, 2024). Second, building on this framework, we conducted an extensive benchmark of twenty VAE variants for pseudo-healthy reconstruction of 3D brain FDG PET (Hassanaly et al., Journal of Medical Imaging, 2025). This benchmark, the first of its kind for this application, established that VAE-based pseudo-healthy reconstruction is feasible even with small training sets (a few hundred subjects), and that the β-VAE is the best overall model in this context.

The PhD of Maëlys Solal (2023–2026) has led to two further contributions. First, we studied two sources of variability in anomaly detection performance: variability in the healthy training population, which can inflate anomaly maps (Solal et al., SPIE Medical Imaging, 2024), and variability across models trained under different random seeds, which makes anomaly detection unreliable and is largely overlooked in the UAD literature (Solal et al., Journal of Medical Imaging, 2026). In both cases, Z-score normalisation of anomaly maps using a healthy control set substantially reduces this variability. Second, in collaboration with Stéphanie Allassonnière (Université Paris Cité) and her PhD student Agathe Senellart, we identified a fundamental trade-off between pseudo-healthy reconstruction quality and anomaly detection performance, governed by the latent space regularisation: less constrained models produce better reconstructions but lower detection performance, while more constrained models improve detection at the cost of reconstruction fidelity (Senellart and Solal et al., ISBI, 2026). We proposed a principled approach to address this trade-off by optimising the latent code at inference time to ensure that the pseudo-healthy reconstruction lies within the healthy distribution while remaining specific to the subject (Solal and Senellart et al., OHBM, 2026).

Finally, Hugues Roy started his PhD in 2024, co-supervised with Reuben Dorent, with a focus on exploring generative architectures beyond VAEs. A first contribution introduced AnoBFN, an extension of Bayesian flow networks to unsupervised anomaly detection, which outperformed state-of-the-art methods based on VAEs, GANs and diffusion models on Alzheimer’s disease-related anomaly detection in FDG PET (Roy et al., DGM@MICCAI, 2025). A second contribution, presented at MICCAI 2026, reformulates unsupervised anomaly detection as a Bayesian inverse problem under a diffusion prior, jointly inferring the pseudo-healthy image and an anomaly mask, and demonstrates improved anomaly localisation compared to other diffusion-based approaches on both FDG PET and anatomical (FLAIR) MRI (Roy et al., MICCAI, 2026).

Main publications

(*) Joint first authorship. (†) Joint last authorship.