Unsupervised out-of-distribution detection for safer robotically guided retinal microsurgery.

Area of Science:

• Machine Learning
• Medical Imaging
• Surgical Robotics

Background:

• Safe machine learning deployment requires identifying out-of-distribution (OoD) samples.
• OoD detection is critical in safety-critical applications like robotically guided retinal microsurgery.
• Instrument-integrated optical coherence tomography (iiOCT) probes acquire 1D images for distance estimation in surgery.

Purpose of the Study:

• Investigate the feasibility of an OoD detector for inappropriate iiOCT images in machine learning-based distance estimation.
• Assess the ability of an OoD detector to identify corrupted samples from real-world ex vivo porcine eyes.

Main Methods:

• Utilized a simple OoD detector based on Mahalanobis distance.
• Evaluated the detector's performance on corrupted iiOCT samples from real-world ex vivo porcine eyes.
• Compared the proposed method (MahaAD) against a supervised approach trained on similar corruptions.

Main Results:

• The Mahalanobis distance-based OoD detector successfully rejected corrupted iiOCT samples.
• The proposed approach maintained downstream task performance within acceptable levels.
• MahaAD outperformed a supervised method and achieved top performance in detecting real-world iiOCT corruptions.

Conclusions:

• Detecting corrupted iiOCT data using OoD detection is feasible and does not require prior knowledge of specific corruptions.
• MahaAD can enhance patient safety in robotic microsurgery by preventing inaccurate distance estimations.
• The method prevents deployed models from making risky predictions on corrupted data.