Brain subtle anomaly detection based on auto-encoders latent space analysis : application to de novo parkinson patients

Nicolas Pinon, Geoffroy Oudoumanessah, Robin Trombetta, Michel Dojat, Florence Forbes, Carole LartizienISBI 2023 · DOI · OpenReview

Introduction

Detecting Parkinson’s Disease (PD) in its earliest stages is notoriously difficult. At the "de novo" stage (just after diagnosis), structural brain changes are often so subtle they remain invisible to the naked eye on standard MRI scans. While supervised deep learning models like CNNs are powerful, they require massive amounts of expertly labeled data, which is rarely available for early-stage pathologies.

Unsupervised Anomaly Detection (UAD) offers a promising alternative by training models to recognize "normal" brain structures. Traditionally, UAD uses Auto-Encoders (AE) and relies on reconstruction error: the assumption that an anomaly cannot be reconstructed well. However, for subtle diseases like early PD, reconstruction error is often too noisy. This paper proposes a more sensitive approach by analyzing the latent space of auto-encoders.

Methods

The researchers developed a framework that shifts the detection step from the final reconstructed image to the compressed mathematical representation, or latent space ($z$).

Schematic of the proposed unsupervised anomaly detection framework, highlighting the transition from reconstruction error (1) to latent space analysis using OC-SVM (2) and MMST (3).
  1. Feature Extraction: A patch-based Siamese Auto-Encoder (SAE) is trained on healthy brain images to capture local spatial context.
  2. Latent Space Analysis: Once trained, latent representations are extracted and analyzed using two multivariate statistical models:
  3. One-Class SVM (OC-SVM): A discriminative algorithm that draws a boundary around healthy data points.
  4. MMST (Mixture of Multiple scale t-distributions): A generative model that accounts for the "heavy tails" in complex biological data to provide a probability score for normality.
  5. Anomaly Mapping: These scores are used to generate maps showing the percentage of abnormal voxels in specific brain regions.

The study utilized the PPMI database, including T1-weighted MRI and Diffusion Tensor Imaging (DTI) from 54 healthy controls and 124 de novo PD patients.

Key Results and Insights

The performance was evaluated using the g-mean score, which balances sensitivity and specificity. The results demonstrated that latent space analysis significantly outperforms traditional methods.

Performance comparison (g-mean) of different detection methods across various brain regions.
  • The Latent Space Advantage: Both OC-SVM and MMST approaches achieved a median g-mean of approximately 0.65 on the whole brain, proving more consistent and sensitive than the reconstruction error baseline.
  • Unsupervised vs. Supervised: In a significant finding, the unsupervised latent space models outperformed supervised 3D ResNet and DenseNet models (which scored between 0.55 and 0.60). This suggests that when disease signals are faint, supervised models may overfit to noise, whereas UAD models benefit from a robust definition of normality.
  • Regional Sensitivity: Anomalies were most detectable in the temporal and occipital lobes. Interestingly, subcortical structures like the substantia nigra showed lower detection performance, suggesting that structural changes there might be too minute for current MRI modalities at this early stage.

Conclusion and Future Work

This research demonstrates that analyzing the latent space of auto-encoders provides a superior mechanism for detecting subtle anomalies in early Parkinson's Disease. While the model currently struggles with deep brain structures, future work will focus on expanding imaging modalities (such as T2-weighted images) and developing hybrid scoring systems that combine reconstruction error with latent space analysis to further improve sensitivity.

Summary updated

Abstract

Neural network-based anomaly detection remains challenging in clinical applications with little or no supervised information and subtle anomalies such as hardly visible brain lesions. Among unsupervised methods, patch-based auto-encoders with their efficient representation power provided by their latent space, have shown good results for visible lesion detection. However, the commonly used reconstruction error criterion may limit their performance when facing less obvious lesions. In this work, we design two alternative detection criteria. They are derived from multivariate analysis and can more directly capture information from latent space representations. Their performance compares favorably with two additional supervised learning methods, on a difficult de novo Parkinson Disease (PD) classification task.
BibTeX
@inproceedings{pinon2023brain,
  author = {Nicolas Pinon and Geoffroy Oudoumanessah and Robin Trombetta and Michel Dojat and Florence Forbes and Carole Lartizien},
  title = {{Brain subtle anomaly detection based on auto-encoders latent space analysis : application to de novo parkinson patients}},
  booktitle = {ISBI},
  year = {2023},
  doi = {10.1109/isbi53787.2023.10230351},
}
Figure