Out-of-Distribution Detection in Medical Image Segmentation with β -VAE and Likelihood Regret.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning model performance degrades with out-of-distribution (OOD) data.
• Existing OOD detection methods primarily focus on image samples or pixels, neglecting OOD segmentation masks.
• Corruptions in ground truth masks are common in real-world medical imaging, necessitating OOD mask detection.

Purpose of the Study:

• To develop a method for simultaneous OOD detection of both image samples and segmentation masks in medical imaging.
• To address the unexamined problem of OOD ground truth annotations in medical image segmentation.
• To enhance the robustness and reliability of deep learning models in clinical settings.

Main Methods:

• A novel model combining beta-Variational Autoencoder (β-VAE) for latent space disentanglement and U-Net for segmentation.
• Introduction of Likelihood Regret to calculate OOD scores for input reconstruction and segmentation, improving distribution fit assessment.
• Simultaneous OOD detection for both image samples and their corresponding segmentation masks.

Main Results:

• The proposed method efficiently discriminates OOD data in 3D medical image segmentation.
• Demonstrated excellent performance in detecting OOD data with large distribution drifts.
• Achieved competitive performance even with very small distribution drifts.

Conclusions:

• The developed method offers a systematic approach to OOD detection in medical image segmentation.
• This work highlights the importance of considering OOD segmentation masks alongside OOD image samples.
• The findings contribute to more reliable deep learning applications in medical imaging.