The main objective of our research program is to create and validate informatics methods for extracting useful content from anatomic, functional and molecular images that will then enable integration of this imaging data with other sources of medically and biologically relevant data.

Quantitation of Molecular Imaging

PET, PET/CT, bioluminescent imaging, fluorescence imaging and other molecular imaging modalities can produce phenomenal images of molecularly-specific activity. However, in scientific studies, these images are often analyzed using semi-quantitative or qualitative measures due to the difficulty in analyzing the content of these images. We are working with several projects within the Molecular Imaging Program at Stanford (MIPS) to help improve the image quantitation and exploratory visualization techniques that are currently used.

Computer Aided Detection of Colonic Polyps in CT Colonography

CT Colonography, or "virtual colonoscopy", produces hundreds of CT images that must be scrutinized to find pre-cancerous polyps which must subsequently be removed. However, this is a very tedious task where human reader performance has been lacking. We have developed computer algorithms that automatically detect potential polyp candidates to present to the human reader. Our evaluations have demonstrated that there is an increase in diagnostic performance when readers are aided by this algorithm. We continue to work on improving the algorithm's detection performance.

We are also working on image warping methods to mathematically unfold the surface of the colon to potentially improve the efficiency of finding polyps. Our initial results are promising and we are continuing work in this area.

Computer Aided Detection of Lung Nodules in Chest CT

High resolution chest CT images must be carefully read in order to detect pre-cancerous lung nodules. This is an exceptionally difficult task to perform unaided. We have developed computer vision algorithms to automatically detect candidate lesions in the lung and present them to a human reader. Initial performance of this algorithm is promising. We continue to work on improving and augmenting this algorithm to improve overall detection performance for lung nodules.