Synthetic 3DCT reconstruction using fluoroscopy and convolutional neural networks for patient-specific real-time image-guided proton therapy
(2024)
Files
THESIS_LOYENEstelle.pdf
Open Access - Adobe PDF
- 19.48 MB
Details
- Authors
- Supervisors
- Macq, Benoit
- Abstract
- External beam radiation therapy is a standard cancer treatment that uses a source of radiation to destroy the tumor. Proton therapy uses a beam of protons to irradiate cancerous tissue. It offers a physical advantage over conventional radiotherapy thanks to the very localised dose deposition of protons within the body. This decreases the risk of side effects because the dose delivered in the surrounding healthy tissue is lower. However, it also means that it is highly vulnerable to uncertainties. A variety of geometrical uncertainties may affect the accuracy of photon and proton therapy, such as respiratory motion, tumor delineation or inter-fraction setup errors. Those inaccuracies are generally overcome by applying safety margins around the target, but larger margins result in increased irradiated healthy tissue. Modern radiation therapy is generally performed using daily image guidance to reduce the uncertainty of overall tumor targeting. However, these technologies are expensive and require the installation of new dedicated devices, not all of which is suitable for proton therapy. The majority of radiation therapy treatment rooms are currently equipped with a projection radiography system. Adaptive radiation therapy is another modern radiation therapy technique that uses imaging information acquired during treatment to re-plan the treatment plan in order to improve target coverage and reduce treatment toxicity. However, the decision to re-plan is made by the radiotherapist and is subject to inter-physician variability. In the context of real-time tumor tracking during treatment delivery, this thesis explores the use of artificial intelligence to reconstruct a 3DCT image from a fluoroscopy image. This research is motivated by the ease of acquiring a x-rays projection in the treatment room, and the need to have a 3DCT image to compute the radiation dose deposition. The reconstructed 3DCT image can be used for several purposes: give a feedback to the machine on the 3D positions of the tumor and internal organs, and/or to compute the radiation dose delivered to the patient. The radiation dose can either be given as feedback to the machine or be used by the radiotherapist to decide whether re-planning is necessary. The research approach taken in this thesis can be divided into three main contributions. The first contribution implements a data augmentation tool to overcome the lack of medical data available to train and validate neural networks. The second contribution focuses on the design of a methodology for reconstructing a 3DCT image from a projection radiography using a patient-specific training of a convolutional neural network. This contribution assesses the quality of the reconstructed images using similarity metrics. The third contribution deals with the use of these images in a proton therapy treatment. To this end, the delivery of a treatment plan on reconstructed 3DCT images is simulated. In each of these last two contributions, a base case and two variants are studied. The aim of the variants is to evaluate and compare the robustness of different training methods to events that may occur in the clinic, such as a change in layout and a change in image acquisition time.
- Affiliations
Citations
Loyen, E. (2024). Synthetic 3DCT reconstruction using fluoroscopy and convolutional neural networks for patient-specific real-time image-guided proton therapy. https://hdl.handle.net/2078.5/259684