Using machine learning to identify critical structures during laparoscopic cholecystectomy
No Preview Available
Citations
Altmetric:
Altmetric::
Download
Authors
File Licence
Editors
Date
2025
Source Title
Journal Title
Journal ISSN
Volume Title
Publisher
Melbourne Authors
Affiliation
Biomedical Engineering
Document Type
PhD thesis
Subject Code
Citation
Access Status
Open Access
DOI
Access URL
Open Access URL
Open Access at PMC
Linked Resource URL
NHMRC Grant Code
ARC Grant Code
Description
© 2025 Yuning Zhou
Abstract
Deep learning has gained significant attention in surgical applications due to its ability to make precise inferences learned from large scale data. In laparoscopic surgery, semantic segmentation plays a crucial role in providing visual support by identifying critical structures. However, training accurate deep learning models remains challenging due to the high cost and time required for large-scale, high-quality annotations.
This thesis addresses this challenge through two key strategies: leveraging self-supervised learning to utilise unlabelled data for improved segmentation performance, and adopting active learning to train deep neural networks in an evolving manner, selecting the most informative data for annotation under constrained budgets.
First, a comprehensive evaluation of SSL methods reveals that pixel-level objectives are the most effective for learning representations in laparoscopic imaging, where anatomy is highly variable and deformable. Based on this insight, a tailored masked denoising autoencoder is proposed, optimising the masking ratio and patch size for surgical segmentation. This method achieves state-of-the-art performance across three laparoscopic datasets, significantly improving segmentation for critical yet under-represented structures while demonstrating strong cross-dataset generalisation ability.
Next, the thesis investigates the adaptation of contrastive learning, a powerful self-supervised learning approach designed for natural images, to surgical data. A novel dimensionality search framework is introduced to automate the selection of optimal data augmentation policies for contrastive learning in a domain-aware manner. It efficiently explores a large search space by leveraging local dimensionality of deep representations as a proxy target, successfully identifying augmentations tailored for laparoscopic surgery. These findings provide valuable insights into domain-specific adaptations required for transitioning contrastive learning to medical applications.
Beyond improving segmentation through unlabelled data, the thesis also explores active learning to enhance model performance under restricted annotation budgets. A dimensionality-aware data querying method is proposed to prioritise informative samples by capturing complexity of geometry in deep representations distributions. While effective, this method alone is insufficient for optimal active learning, as it does not explicitly account for model uncertainty or data diversity. Future work should explore hybrid approaches to balance distributional complexity, model confidence, and clinical relevance.
By addressing both unlabelled data utilisation and annotation efficiency, this thesis contributes novel methodologies that advance data-efficient deep learning in surgical applications, paving the way for more accurate and scalable segmentation models in real-world clinical settings.
Notes
Terms of Use
Keywords
Deep learning, Self-supervised learning, Representation learning, Semantic segmentation, Active learning, Medical image analysis, Laparoscopic surgery, Computer-assisted surgery