Image-conditioned latent rectified flow models for 3D medical anomaly localisation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Unsupervised anomaly detection in medical imaging often relies on reconstruction-based methods.
• These methods struggle with assumptions about healthy tissue reconstruction and pathology reproduction.
• Existing 2D diffusion models for anomaly detection are computationally expensive and ignore volumetric context.

Purpose of the Study:

• To develop a more effective and efficient unsupervised anomaly detection method for medical imaging.
• To address limitations of 2D pixel-space methods by incorporating 3D volumetric context.
• To reduce computational overhead while maintaining or improving detection performance.

Main Methods:

• Implemented image-conditioned restoration in a 3D latent space using a Variational Autoencoder (VAE) and rectified flow.
• Proposed a "restoration change" metric to mitigate false positives from VAE compression.
• Experimented with training anomaly detection tasks directly in latent space for improved sensitivity.

Main Results:

• Achieved state-of-the-art performance on medical imaging benchmarks, including brain MRI and the AADD dataset.
• Demonstrated a significant reduction in computational cost (approximately 10x faster) by using 3D latent space.
• An ensemble of models trained with both pixel-space and latent-space anomalies showed the strongest results.

Conclusions:

• Incorporating 3D context in latent space significantly enhances unsupervised anomaly detection performance and efficiency.
• The proposed "restoration change" metric effectively reduces false positives.
• Further research into anomaly imputation strategies can lead to more robust models for challenging benchmarks like AADD.