Atzeni, Alessia; (2023) Semi-automated learning strategies for large-scale segmentation of histology and other big bioimaging stacks and volumes. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text PhD_Alessia_Atzeni.pdf - Accepted Version Download (75MB) | Preview

Abstract

Labelled high-resolution datasets are becoming increasingly common and necessary in different areas of biomedical imaging. Examples include: serial histology and ex-vivo MRI for atlas building, OCT for studying the human brain, and micro X-ray for tissue engineering. Labelling such datasets, typically, requires manual delineation of a very detailed set of regions of interest on a large number of sections or slices. This process is tedious, time-consuming, not reproducible and rather inefficient due to the high similarity of adjacent sections. In this thesis, I explore the potential of a semi-automated slice level segmentation framework and a suggestive region level framework which aim to speed up the segmentation process of big bioimaging datasets. The thesis includes two well validated, published, and widely used novel methods and one algorithm which did not yield an improvement compared to the current state-of the-art. The slice-wise method, SmartInterpol, consists of a probabilistic model for semi-automated segmentation of stacks of 2D images, in which the user manually labels a sparse set of sections (e.g., one every n sections), and lets the algorithm complete the segmentation for other sections automatically. The proposed model integrates in a principled manner two families of segmentation techniques that have been very successful in brain imaging: multi-atlas segmentation and convolutional neural networks. Labelling every structure on a sparse set of slices is not necessarily optimal, therefore I also introduce a region level active learning framework which requires the labeller to annotate one region of interest on one slice at the time. The framework exploits partial annotations, weak supervision, and realistic estimates of class and section-specific annotation effort in order to greatly reduce the time it takes to produce accurate segmentations for large histological datasets. Although both frameworks have been created targeting histological datasets, they have been successfully applied to other big bioimaging datasets, reducing labelling effort by up to 60−70% without compromising accuracy.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as XMLJSONCSV

Archive Staff Only

View Item