Self-supervised out-of-distribution detection in wireless capsule endoscopy images.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning models excel in many areas but struggle with out-of-distribution (OOD) inputs, which are critical in medical applications for detecting rare diseases or anomalies.
• Robust detection of OOD medical images is essential for patient safety and accurate diagnosis.

Purpose of the Study:

• To develop a novel patch-based, self-supervised approach for improved OOD detection in wireless capsule endoscopy (WCE) images.
• To create a system capable of identifying unseen pathologies and anomalies without relying on labeled datasets.

Main Methods:

• A three-stage self-supervised method was employed, starting with training a triplet network for WCE image patch representation learning.
• Patch embeddings were clustered based on visual similarity, and these cluster assignments were used as pseudolabels.
• A patch classifier was trained using pseudolabels, incorporating the Out-of-Distribution Detector for Neural Networks (ODIN) for OOD detection.

Main Results:

• The proposed method demonstrated improved OOD detection performance on the Kvasir-capsule WCE dataset compared to baseline approaches.
• The system successfully detected unseen pathologies and anomalies, including lymphangiectasia, foreign bodies, and blood, achieving an Area Under the Receiver Operating Characteristic curve (AUROC) greater than 0.6.
• The approach proved effective in OOD detection without the need for labeled images.

Conclusions:

• This work presents an effective and novel self-supervised solution for OOD detection in medical imaging, specifically WCE.
• The patch-based approach enhances the robustness of deep learning models in identifying anomalous or rare findings.
• The method offers a valuable tool for medical diagnostics, particularly in scenarios with limited or no labeled data for rare conditions.