Limitations of Out-of-Distribution Detection in 3D Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models exhibit performance degradation on data from distributions different from their training data.
• Out-of-distribution (OOD) detection is crucial in critical applications like medical imaging to prevent erroneous predictions.
• Existing OOD detection methods have not been thoroughly evaluated for 3D medical image segmentation tasks.

Purpose of the Study:

• To investigate the effectiveness of OOD detection methods in the context of 3D medical image segmentation.
• To identify limitations of current OOD detection techniques when applied to 3D medical imaging.
• To propose a novel, effective baseline method for OOD detection in this domain.

Main Methods:

• Designed several out-of-distribution (OOD) challenges simulating clinically relevant scenarios for 3D medical image segmentation.
• Evaluated general OOD detection methods and methods specifically designed for segmentation.
• Developed and tested a novel method, Intensity Histogram Features (IHF), based on intensity distributions.

Main Results:

• None of the evaluated OOD detection methods achieved acceptable performance on the designed 3D medical image segmentation challenges.
• General OOD methods performed poorly, with the best mean false-positive rate (FPR) at 95% true-positive rate (TPR) of 0.59.
• Segmentation-dedicated methods showed suboptimal performance (best mean FPR: 0.31), while the proposed IHF method achieved comparable or better results (mean FPR: 0.25).

Conclusions:

• Existing OOD detection methods have significant limitations for 3D medical image segmentation.
• The proposed Intensity Histogram Features (IHF) method offers a promising and effective baseline for OOD detection in this field.
• The study releases a benchmark dataset and evaluation criteria to advance research in OOD detection for 3D medical imaging.