Affiliation: This project was done with Zack Markow and Richard Pham for our ESE 589, Biological Imaging Technology, project.
Project Description: We were interested in reducing the radiation dose of early detection mammography for breast cancer. The frequency and timing of these scans are chosen as a balance of probability between finding an actual tumor and needlessly irradiating the patient. If we could come up with a way to perform mammographies with lower radiation dose, doctors could prescribe them more often, catching more tumors early on with less side effects. Additionally, we hoped that the algorithms we use here can be adapted for use in X-ray CT where radiation dose is a much more severe issue.
We ended up designing an adaptive sensing modality. Due to Poisson sampling noise, x-ray image resolution is proportional to the square root of the number of photons. While a doctor is only going to be interested in getting high resolution around possible tumors, the X-ray machine is indiscriminate and blasts the entire breast with equal radiation. Our idea was to first do a low-dose, low-resolution scan and use machine learning to identify regions likely to contain a tumor. Additional radiation would then be applied to only those areas, giving the doctor the resolution needed to make a diagnosis without irradiating the entire breast. See below for our proposal