In the computational medical imaging and therapy lab, we develop imaging science, machine-learning and physics-based computational imaging algorithms such that these algorithms yield optimal performance on the underlying clinical tasks. For example, in an image of a patient with cancer symptoms, the task could be detecting the tumor or estimating some tumor property such as volume. Thus, the designed imaging algorithm should yield the best performance in that task. This task performance can be objectively quantified using signal processing and computational imaging methods. This idea of objective assessment of image quality (OAIQ) for designing medical imaging solutions is the overarching theme of research in our lab.
Following are some current research directions:
- LC-QSPECT: Ultra low-count quantitative SPECT for guiding treatments: We are developing new low-count quantitative imaging methods to estimate regional tracer uptake for alpha-particle radiopharmaceutical therapies. A key challenge in this quantification task is the ultra low number of detected counts (20 times lower than conventional SPECT). We are developing approaches that directly estimate uptake from the projection data and skip the reconstruction step. These reconstruction-less estimation methods are observed to yield accurate performance on quantification tasks. We are also developing methods that use list-mode data to advance quantitative SPECT. This research direction is supported by an NIH R01 award.
- Task-specific design of medical imaging algorithms: We develop techniques to optimize and evaluate imaging systems on specific clinical tasks. A specific area of interest is no-gold-standard evaluation (details below). We also collaborate with the wider community towards these efforts. For example, the PI recently led multi-institutional efforts to provide best practices for evaluation of AI-based methods (RELAINCE guidelines).
- No-gold-standard evaluation: Evaluating quantitative imaging methods with patient data: A specific effort in the lab has been on the evaluation of quantitative imaging methods with patient data. Such evaluation typically requires the availability of a gold standard, but these are usually unavailable. To address this issue, we are developing no-gold-standard evaluation techniques, that, without the availability of a gold standard or the true quantitative value, are able to evaluate quantitative imaging methods. This project is supported by an NIH R01 award
- Deep-learning-based quantitative PET methods for early prediction of therapy response in lung cancer: We are developing novel machine-learning-based methods to segment tumors and quantify volumetric and radiomic features from PET images and evaluating these features as biomarkers for predicting therapy response in patients with cancer. The work is supported by an NIH R56 award.
- Physics and deep-learning-based CT-less SPECT: We are developing methods to enable transmission-less attenuation compensation in cardiac SPECT, with the goal of helping reduce dose and costs, improving diagnosis, and increasing patient comfort. This project was supported by the NIBIB Trailblazer award.
- Deep-learning-based methods for brain SPECT: We are developing new image reconstruction and analysis methods with the goal of developing SPECT-based biomarkers to measure severity of Parkinson disease.