Colleoni, Emanuele; (2025) Towards surgical image segmentation supported by synthetic data. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text PhD_Thesis_Colleoni.pdf - Submitted Version Download (177MB) | Preview

Abstract

Robotic minimally invasive surgery was developed to overcome key challenges of traditional minimally invasive procedures, such as restricted access and counterintuitive instrument control. By offering enhanced dexterity, 3D vision, and tremor mitigation, robotic systems allow surgeons to operate with greater ease and precision. These platforms also provide a strong foundation for learning-based technologies, thanks to digital video streams and rich kinematic data from robotic arms. This enables AI algorithms that assist surgeons, for example, by guiding navigation or recognizing anatomical and pathological structures. Image understanding methods, particularly anatomy and tool segmentation, are central to developing and deploying these technologies. However, training such models requires large amounts of annotated data, which is costly, time-consuming, and demands expert knowledge. Without sufficient high-quality data, models often underperform and produce unreliable results. This shortage remains a major challenge for surgical AI. This thesis addresses this bottleneck by proposing surgical data generation techniques to reduce reliance on manual annotation for anatomy and instrument segmentation. It first introduces a pipeline for generating annotated segmentation videos and proposes a deep learning model that combines video and simulation data, improving performance in challenging scenarios with blood or cauterising smoke compared to standard models. To further address data scarcity, the thesis proposes data generation pipelines that combine image-to-image translation with surgical simulators and introduce strategies to enhance image quality and realism. Generative models integrating recent architectural advances are proposed and evaluated across multiple surgical datasets. Results show that synthetic data can train anatomy and instrument segmentation models to perform comparably to those trained on real data, and that combining real and synthetic data further improves segmentation performance in standard pipelines. Overall, the thesis advances efforts to reduce segmentation models’ dependence on limited surgical data by proposing and validating artificial data generation techniques that improve training and performance.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item