Abstract
This thesis presents the development and evaluation of three deep learningmodels to address the need for quantification in 3D imaging. With advancements in Computed Tomography (CT) and confocal fluorescent microscopy
producing high-resolution 3D images, these models aim to generalize across
heterogeneous datasets. The first model tracks submicron colloids using confocal nanoscopy, improving predictions in high-density and low-contrast environments while being easily generalized to multiple suspensions. The second
model is for zebrafish otolith semantic segmentation from µCT images at 0.83
Dice accuracy, while the third model automatically segments lower jaws from
µCT images at 0.73 Dice accuracy. The overall findings emphasize the criticality of data cleaning, simulation, and 3D models for accurate segmentation.
Additionally, models can be trained effectively even with small datasets, and
it is possible to focus on either precision or recall, depending on the task requirements. These deep learning models show promising results for increasing
accuracy and reducing subjectivity for image quantification in three dimensions.
Date of Award | 3 Oct 2023 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | Chrissy L Hammond (Supervisor) & Danielle M Paul (Supervisor) |