Unsupervised domain adaptation for clinician pose estimation and instance segmentation in the operating room

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Accurate clinician localization in operating rooms (ORs) is crucial for advanced OR support systems.
• Existing computer vision models struggle with OR-specific image characteristics and privacy concerns, hindering data collection and annotation.
• There's a need for robust models that can handle domain shift and limited labeled data in OR environments.

Purpose of the Study:

• To investigate joint person pose estimation and instance segmentation on low-resolution OR images.
• To develop an unsupervised domain adaptation method (AdaptOR) for adapting models from general datasets to specific OR domains without labeled OR data.
• To address privacy concerns by enabling effective analysis on low-resolution, privacy-preserving images.

Main Methods:

• Studied joint person pose estimation and instance segmentation on images downsampled up to 12x.
• Proposed AdaptOR, an unsupervised domain adaptation technique using geometric constraints and pseudo-labeling for self-training on unlabeled OR data.
• Introduced disentangled feature normalization to manage domain discrepancies between source and target datasets.

Main Results:

• AdaptOR demonstrated effectiveness in adapting models to low-resolution, privacy-preserving OR images, outperforming baseline methods.
• The approach showed strong performance on MVOR+ and TUM-OR-test datasets, particularly with reduced image resolution.
• As a semi-supervised learning method on the COCO dataset, it achieved comparable results to fully supervised models using only 1% of labeled data.

Conclusions:

• AdaptOR effectively addresses domain shift and data scarcity challenges in OR computer vision.
• The method enables privacy-preserving clinician localization using low-resolution images, crucial for developing next-generation OR support systems.
• AdaptOR shows versatility, performing well in both unsupervised domain adaptation and semi-supervised learning scenarios.