edoc

Automated Distinct Bone Segmentation from Computed Tomography Images using Deep Learning

Schnider, Eva. Automated Distinct Bone Segmentation from Computed Tomography Images using Deep Learning. 2023, Doctoral Thesis, University of Basel, Faculty of Medicine.

[img]
Preview
PDF
16Mb

Official URL: https://edoc.unibas.ch/93876/

Downloads: Statistics Overview

Abstract

Large-scale CT scans are frequently performed for forensic and diagnostic purposes, to plan and
direct surgical procedures, and to track the development of bone-related diseases. This often
involves radiologists who have to annotate bones manually or in a semi-automatic way, which is
a time consuming task. Their annotation workload can be reduced by automated segmentation
and detection of individual bones. This automation of distinct bone segmentation not only has
the potential to accelerate current workflows but also opens up new possibilities for processing
and presenting medical data for planning, navigation, and education.
In this thesis, we explored the use of deep learning for automating the segmentation of all
individual bones within an upper-body CT scan. To do so, we had to find a network architec-
ture that provides a good trade-off between the problem’s high computational demands and the
results’ accuracy. After finding a baseline method and having enlarged the dataset, we set out
to eliminate the most prevalent types of error. To do so, we introduced an novel method called
binary-prediction-enhanced multi-class (BEM) inference, separating the task into two: Distin-
guishing bone from non-bone is conducted separately from identifying the individual bones.
Both predictions are then merged, which leads to superior results. Another type of error is tack-
led by our developed architecture, the Sneaky-Net, which receives additional inputs with larger
fields of view but at a smaller resolution. We can thus sneak more extensive areas of the input
into the network while keeping the growth of additional pixels in check.
Overall, we present a deep-learning-based method that reliably segments most of the over
one hundred distinct bones present in upper-body CT scans in an end-to-end trained matter
quickly enough to be used in interactive software. Our algorithm has been included in our
groups virtual reality medical image visualisation software SpectoVR with the plan to be used
as one of the puzzle piece in surgical planning and navigation, as well as in the education of
future doctors.
Advisors:Cattin, Philippe Claude
Committee Members:Jost, Gregory and Trägårdh, Elin and Huck , Antal
Faculties and Departments:03 Faculty of Medicine > Departement Biomedical Engineering > Imaging and Computational Modelling > Center for medical Image Analysis & Navigation (Cattin)
UniBasel Contributors:Cattin, Philippe Claude
Item Type:Thesis
Thesis Subtype:Doctoral Thesis
Thesis no:14994
Thesis status:Complete
Number of Pages:xiv, 113
Language:English
Identification Number:
  • urn: urn:nbn:ch:bel-bau-diss149946
edoc DOI:
Last Modified:21 Apr 2023 04:30
Deposited On:20 Apr 2023 09:49

Repository Staff Only: item control page