UC Irvine

Radiation - a true double-edged sword of healthcare. At once, a powerful toxin, capable of causing irreversible damage to every system of the body and death at sufficient doses. Yet all the while, a cornerstone of cancer treatment, having been utilized in the treatment of 50% of all cancer patients for decades. Despite its extensive use, the damage radiotherapy causes to healthy tissue remains a significant impact on quality of life and the greatest dose-limiting factor. Therefore, there exists an urgent need for effective methods to mitigate radiation-induced injury as well as reliable non-invasive monitoring techniques. A novel technology that can significantly reduce injury to healthy tissue is FLASH radiotherapy - delivering equivalent doses of radiation at ultra-high dose rates. While current dose rates (referred to as conventional or CONV) range from 0.07-0.1 Gy/s, FLASH radiotherapy (FLASH-RT) typically exceeds 100 Gy/s. It was discovered that when equivalent doses and fractionation schedules are delivered with FLASH or CONV, tumor killing ability was identical, but FLASH-RT significantly spared the damage to healthy tissue and was often indistinguishable from unirradiated controls. In the years since its rediscovery, investigators have studied the boundaries and possible mechanisms of FLASH. As of 2024 it has been demonstrated in electrons, protons and photons at a range of doses and fractionation paradigms. FLASH sparing has been effective in zebrafish, rodents, cats, and humans and can prevent sequelae to the lungs, skin, immune system, heart and brain. While several mechanistic explanations for the FLASH effect have been hypothesized and evaluated, none are yet to be successful in answering the question of why FLASH really works. An assessment technique that has prominent use in radiation oncology, but to date has not been employed to study FLASH, is medical imaging. Medical images are powerful diagnostic tools that have provided non-invasive methods of monitoring patient health since their invention. However, acquisition of these images is often expensive, time-intensive, and can also contribute to additional radiation exposure. Therefore, there is an incentive to minimize the number of images acquired while maximizing the information output for each image. To this end, artificial intelligence is entering the frame. Artificial Intelligence (AI) is revolutionizing healthcare in a number of ways, exploiting the vast sets of patient health data to search for patterns that are otherwise undetectable. Already AI tools are being employed to automate workflows, segment tumors and even predict treatment outcomes. Medical images are an obvious sandpit for AI to play in, given the large number that are generated globally and the vast amount of data that can be extracted from a single image. The burgeoning field of radiomics is borne from this idea, that if a picture is worth a thousand words, an image is worth thousands of datapoints. In a typical radiomics workflow, a region of interest is segmented from each image, and then hundreds of predefined features are extracted from the region. These features can then be compared to clinical and biological endpoints or be used to train machine learning algorithms. Radiomic features are yet to be tied to biological characteristics but have the potential to extract information from images in a repeatable and examinable manner. Deep learning can also extract features from medical images but is received with more hesitation by healthcare workers because of its “black box” nature. While advancements are occurring at an exponential rate within each of these domains (radiotherapy, medical imaging and artificial intelligence), these spheres are yet to fully collide. In this thesis I describe studies that touch on all of these disciplines for the first time. Firstly, an investigation of the reproducibility and temporality of the FLASH effect on the brain is presented, demonstrating that FLASH sparing can be reproduced between institutions and that the early response to FLASH does not differ from the early response to CONV dose-rate radiation. This study was followed up by a Magnetic Resonance Imaging (MRI) study of cranial FLASH-RT. The sparing of cognitive function by FLASH was replicated and changes in the structural and diffusion-weighted MR images were observed in the CONV cohort that was reduced or avoided in the FLASH cohort. In addition, a preliminary radiomics analysis was conducted on the structural images that indicated a stronger discrimination ability of machine learning models to distinguish between controls and CONV-RT animals and a weaker discrimination ability between controls and FLASH-RT animals. Lastly, a deeper dive into the potential of radiomics to identify radiation-induced damage is presented in a cohort of thoracically irradiated mice with and without extra-cellular vesicle treatment. While radiation pneumonitis was not observed through traditional methods, radiomic features were able to detect radiation differences and predict radiation status with a machine learning model. My dissertation research utilizes machine learning and other analytical methods in a novel application, to probe the impact of ionizing radiation on the brain and inspect the FLASH effect in a way it has never been studied before.