Psychogyios, Dimitrios; (2025) Deep Learning Architectures for Disparity Estimation and Segmentation in Surgery. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text d_psychogyios_thesis.pdf - Accepted Version Download (53MB) | Preview

Abstract

The introduction of robotics in the operating room has addressed significant limitations of minimally invasive surgery. By incorporating stereo endoscopic cameras, 3D displays, and intuitive control interfaces, robotic surgery systems provide surgeons with tools to enable more precise operations, while allowing the integration of computer-assisted intervention pipelines. A detailed 3D representation of the surgical environment is essential for existing computer-assisted intervention systems and holds promise for the advancement of surgical automation. This thesis focuses on the development of learning-based methods for estimating 3D information from stereo-endoscopic videos and addresses challenges related to data generation and stereo-matching approaches. To enhance data-driven algorithms that depend on high-quality training data, this thesis introduces a new stereoscopic dataset tailored for surgical environments, facilitating the evaluation and adaptation of 3D reconstruction techniques. Building on this foundation, a multi-task learning framework was developed for the simultaneous estimation of stereo disparity and surgical tool segmentation. This framework leverages shared network architectures to mitigate data scarcity and optimize inference time. Validation studies demonstrated that cross-task learning can enhance performance, enabling the model to adapt its disparity estimation capabilities using only monocular segmentation data, thus illustrating domain transferability. In addition, the thesis explores the application of neural rendering techniques through the development of realistic endoscopic illumination models based on Neural Radiance Fields. The introduced method extends conventional neural radiance field architectures to accurately model the unique illumination conditions found in surgical settings, enhancing the realism and fidelity of synthetic data generation. The integration of this illumination modelling technique facilitates the generation of realistic looking training datasets with low-noise depth, and pose data, contributing to the improvement of learning-based 3D reconstruction and depth estimation in surgical applications. The results highlight that such realistic synthetic data can be used for training and evaluation of deep learning models in surgery, ultimately helping to reduce the domain gap for machine learning models in surgical environments.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item