Machine learning approaches to image processing for alpha-particle radiopharmaceutical microdosimetry
Youngho Seo, Professor
UC San Francisco
Closed. This professor is continuing with Spring 2024 apprentices on this project; no new apprentices needed for Fall 2024.
Keywords: cancer, radiopharmaceutical therapy, targeted alpha therapy, alpha particles, image processing, dosimetry, machine learning, digital autoradiography
Radiopharmaceutical therapy with alpha-particle emitters (⍺RPT) is an emerging cancer treatment method that has demonstrated high efficacy in clinical trials for several types of cancer. Improvement of these drugs requires thorough pre-clinical and clinical dosimetry—the study of the agent’s localization and the resulting deposition of ionizing radiation in cancerous and healthy tissues. ⍺-particle emitters pose unique radiation detection and dosimetry challenges because their short range in tissue (<100 µm) and high energy deposition results in low administered activities and low measurable signals in vivo (e.g. from gamma-rays).
To circumvent these challenges, we use digital autoradiography to directly measure ⍺-particles from ex vivo tissue samples. Digital autoradiographs use thin tissue samples, similar to microscope slides, in conjunction with a scintillator and high-resolution CCD or CMOS camera to obtain high-resolution ⍺-particle emission maps.
Our team uses these radioactivity maps to assess the radiation dose in tissues. Further analysis with Monte Carlo simulation, corroboration with immunohistological stains (e.g. hematoxylin & eosin), or statistical analysis is typical. These all require image processing, including registration (alignment) of consecutive image slices, pixelization or interpolation of event locations, and segmentation of cell nuclei in H&E images, among others.
The accuracy of dose measured from this workflow is directly dependent on the accuracy of these steps. The student will investigate machine learning approaches to improve upon or automate these processes, in which only basic brute-force and often manual methods are currently implemented.
Role: Training Opportunity
- Understand image processing as a requirement for image-based data analysis
- Research available literature and select one or more steps as candidates for improvement (segmentation, alignment, etc)
- Train or implement a pre-trained machine learning model for the task and test it on real data collected from radioactive tissue samples
- Integrate the task with a larger dosimetry calculation workflow
Qualifications: (Required) Some familiarity with programming languages, Python preferred; (Preferred) Some familiarity with machine learning models or image processing; (Bonus) Experience with any of the following: nuclear physics, radiation detection, microscopy, immunohistology, or related fields
Day-to-day supervisor for this project: Robin Peter, Graduate Student
Hours: 6-8 hrs
Off-Campus Research Site: Programming work can be done remotely. The project mentor is available to meet online or in person at UC Berkeley or UCSF campus for meetings and discussions.
Related website: http://www.radiology.ucsf.edu/physics
Related website: https://doi.org/10.1038/s41598-022-22664-5