InfoOOD: information bottleneck optimization for post hoc medical image out-of-distribution detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Information Theory

Background:

• Deep learning models struggle with out-of-distribution (OOD) data, leading to potential failures in clinical applications.
• Existing OOD detection methods lack sensitivity to subtle variations common in medical imaging.

Purpose of the Study:

• To introduce and validate InfoOOD, a novel post hoc, information-based approach for detecting OOD data in medical images.
• To assess InfoOOD's sensitivity and clinical relevance compared to existing OOD detection methods.

Main Methods:

• Trained a 3D U-Net for liver and lesion segmentation on abdominal CT images (N=157).
• Simulated physics-based artifacts (low dose, sparse view, ring artifacts) on test images (N=40).
• Evaluated InfoOOD's detection performance against embedded feature-based and reconstruction-based methods.

Main Results:

• Artifact simulation significantly degraded segmentation performance, worsening with artifact magnitude.
• InfoOOD consistently outperformed existing methods in detecting artifact-induced OOD data (e.g., AUC=0.93 vs. 0.57 for strong ring artifacts).
• InfoOOD showed stronger negative correlations with segmentation performance metrics, indicating better reliability.

Conclusions:

• InfoOOD is a novel, highly sensitive, and clinically relevant method for medical image OOD detection.
• This approach supports the safe deployment of deep learning models in clinical environments by reliably identifying data shifts.