University of Twente Student Theses


Towards Longitudinal Three-Dimensional Photoacoustic Imaging: An Automatic Image Registration Framework

Kim, BSc. N.Y. (2023) Towards Longitudinal Three-Dimensional Photoacoustic Imaging: An Automatic Image Registration Framework.

[img] PDF
Abstract:Problem statement: According to Global Cancer Statistics, breast cancer was the most diagnosed cancer worldwide in 2020, which caused 6.9% of the total cancer deaths, leading to a substantial societal and economic impact. Neoadjuvant chemotherapy (NAC) is a systemic treatment before tumor surgery and it has made significant progress in the overall survival rate and is becoming standard care in breast cancer. However, not every patient responds to NAC. In such cases, a whole course of NAC can lead to a delayed start of another treatment and unnecessary exposure of the patients to the toxic effects of the drugs. Therefore, a longitudinal imaging protocol is needed to acquire multiple images over time to monitor the effects of NAC and prevent further disease progression due to delays in treatment. Ultrasound and magnetic resonance imaging are common clinical imaging modalities primarily employed for structural imaging., However, these modalities have several drawbacks. Their main limitation is that these modalities detect structural changes, such as a change in tumor size, which occur with a delay to changes in the tumor microstructure and do not correlate with patient outcomes. Photoacoustic (PA) imaging is one of the most emerging modalities for capturing and quantifying tumor angiogenesis and hypoxia. It can visualize blood vessels in the breast with sub-millimeter resolution at depths of more than 5 cm. However, a valid image registration method is needed to align PA images over time to monitor NAC's effect. Repositioning the breast leads to complex and non-linear deformations of the vasculature and the breast, making quantitative analysis impossible. Current image registration frameworks are not suitable for image registration of PA images due to limitations like requiring a large training data set or not having functionalities to align sparse data, like PA images. Aim and approach: This thesis proposes a new robust machine learning framework, MUVINN, which uses a coordinate-based neural network to represent the displacement field of the PA image pair. By using a loss function based on normalized cross-correlation and Frangi vesselness filter at multiple scales, it can align vascular images effectively. The algorithm is tested on an in vivo data set of breast PA images of a healthy volunteer acquired with the Twente Photoacoustic Mammoscope 3 to validate the framework. First by synthetically deforming the existent images and then by repeating measurements after repositioning the breast under normal conditions and unfavorable conditions, such as using different illumination wavelengths, purposefully mispositioning the volunteer, using a different breast-supporting cup size, and without the use of a cup. Results: MUVINN has shown excellent performance in registering synthetically deformed images and repeated images in normal conditions and challenging conditions. It has been shown to be robust to shifts in image intensity and field-of-view inconsistencies. Conclusions: MUVINN is a promising tool for quantitatively monitoring disease progression and treatment response in breast cancer using photoacoustic. It has been proven to work for unimodal image registration of PA images with a Twente Photoacoustic Mammoscope 3 imager. However, more research is needed for multimodal image registration for broader applications.
Item Type:Essay (Master)
Faculty:TNW: Science and Technology
Subject:44 medicine, 50 technical science in general
Programme:Biomedical Engineering MSc (66226)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page