Thesis
Hardware acceleration using FPGAs for adaptive radiotherapy
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2021
- Thesis identifier
- T15912
- Person Identifier (Local)
- 201390505
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- Adaptive radiotherapy (ART) seeks to improve the accuracy of radiotherapy by adapting the treatment based on up-to-date images of the patient's anatomy captured at the time of treatment delivery. The amount of image data, combined with the clinical time requirements for ART, necessitates automatic image analysis to adapt the treatment plan. Currently, the computational effort of the image processing and plan adaptation means they cannot be completed in a clinically acceptable timeframe. This thesis aims to investigate the use of hardware acceleration on Field Programmable Gate Arrays (FPGAs) to accelerate algorithms for segmenting bony anatomy in Computed Tomography (CT) scans, to reduce the plan adaptation time for ART. An assessment was made of the overhead incurred by transferring image data to an FPGA-based hardware accelerator using the industry-standard DICOM protocol over an Ethernet connection. The rate was found to be likely to limit the performanceof hardware accelerators for ART, highlighting the need for an alternative method of integrating hardware accelerators with existing radiotherapy equipment. A clinically-validated segmentation algorithm was adapted for implementation in hardware. This was shown to process three-dimensional CT images up to 13.81 times faster than the original software implementation. The segmentations produced by the two implementations showed strong agreement. Modifications to the hardware implementation were proposed for segmenting fourdimensional CT scans. This was shown to process image volumes 14.96 times faster than the original software implementation, and the segmentations produced by the two implementations showed strong agreement in most cases.A second, novel, method for segmenting four-dimensional CT data was also proposed. The hardware implementation executed 1.95 times faster than the software implementation. However, the algorithm was found to be unsuitable for the global segmentation task examined here, although it may be suitable as a refining segmentation in the context of a larger ART algorithm.
- Advisor / supervisor
- Nailon, Bill
- Crockett, Louise H. (Louise Helen)
- Resource Type
- DOI
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
PDF of Thesis T15912 | 2021-09-06 | Public | Download |