Every Click Counts: Deep Learning Models for Interactive Segmentation in Biomedical Imaging
Name
Donatas Vaiciukevičius
Abstract
The medical field of radiology is experiencing an unprecedented surge in demand, largely driven by an ageing population and consequent pressures on an already understaffed workforce. These pressures have resulted in an increased demand for technological advancements to manage the escalating workload. Although various machine learning solutions have been implemented in some areas to alleviate this strain, significant challenges remain. Notably, manually measuring tumours detected in computed tomography images, a key task in the oncological diagnostic workflow, continues to be labour-intensive and presents an opportunity for improvement. To this end, in this thesis, we explored the potential of employing interactive deep-learning models to assist radiologists, thereby improving diagnostic workflow. We experimented with techniques, such as RITM and FocalClick, for the analysis of computed tomography images. This analysis facilitated our primary contribution - the introduction of dynamic radius disk encoding, which adds a new dimension to the click-based user input and substantially boosts model performance. This innovation reduces the need for repeated interactions and enhances segmentation quality with fewer clicks. Additionally, we present an improved augmentation strategy and introduce a novel metric for evaluating interactive segmentation models. Our results demonstrate the effectiveness of these deep learning methods, particularly when augmented with dynamic radius disk encoding, in streamlining radiological diagnostics. The findings represent a promising avenue for further research aimed at further optimising these techniques for clinical application.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Dmytro Fishman
Defence year
2024