Image data science
Our laboratory develops computational predictive models for applications in pathology, radiology, systems biology, as well as mobile and remote sensing. Our mathematical models are based on hierarchical, multi-scale, physics and physiology inspired representation techniques and, in combination with statistical machine learning techniques, are used to regress structure-function relationships at the cellular, tissue, and organ scales. We’re particularly interested in designing low-level processing techniques (e.g. segmentation, registration, tracking), as well as high-level data analytics methods (e.g. detection, estimation, classification) to jointly mine signal, image, molecular, and genomic data to better understand differences between healthy versus diseased states.
Current projects include the development of mathematical signal and image transforms and hierarchical models for inference and predictive modeling. Applications currently being investigated include understanding cell function and morphology for cancer diagnosis and prognosis, brain morphology and function for analyzing healthy vs diseased populations, as well as detection, classification, and inverse problems in mobile and remote sensing.
- Signal transforms and machine learning: telling signal and patterns apart is at the heart of many modern problems in science and technology. We are developing new invertible signal transforms, with well defined forward (analysis) and inverse (synthesis) operations, for machine learning and pattern recognition. Preliminary results have shown that the signal transformation framework can significantly augment the classification performance in detection problems (e.g. cancer detection from histopathology images, face identification, etc) in comparison to state of the art methods. The new signal transformation framework is generative, and thus allows for modeling and visualization of intensity variations in a signal (image) database. The approach can be used to `visualize' any decision (classifier) boundary computed in transform space. It is completely automated, does not require the determination or existance of corresponding landmarks, and consitutes a signal transform with analysis and synthesis operations well defined. Click here for more information.
- Predictive modeling in cancer: enabled by the signal/image transformation framework described here we are in the process of constructing predictive models for cancer research and clinical applications. The idea is to combine image analytics methods to extract meaningful predictors from microscopy image data that, in combination with genetic and molecular data, can recapitulate grading, diagnostic, prognostic and theragnostic information. We have utilized histopathology and cytology images to build diagnostic systems for lung, thyroid, liver, and skin cancers. We are utilizing hierarchical methods, in combination with image analytis, to build automatic, end-to-end systems for comprehensive analysis of pathology data. Click here for more information.
- Image-based cytometry and cell modeling: Modern imaging techniques are able to measure information regarding cellular processes with increasing accuracy, and specificity. Numerous applications in health sciences (drug discovery, genetic screens, diagnosis, prognosis, etc.) can be benefited by image data analysis techniques capable of deriving relevant biological information from such datasets. We are developing new approaches for mining information contained in cell image databases, and utilizing it to model important cellular processes. The approach draws upon our contributions to deformation and transport-based morphometry, as well as other parametric and non parametric extensions. Our efforts are summarized in the CellOrganizer project webpage.
- Biomedical image analysis : We have developed a series of algorithms for image registration for quantitative applications in biomedicine. Our multimodal nonrigid registration methods have been used for template-based image segmentation, population analysis, motion correction, distortion correction, and other applications. The methods work well with multi-modal data, since they are based on mutual information optimization. We have also investigated interpolation artifacts, and their effect on image registration itself, as well as their effect on quantitative estimates produced based on registered data. In the papers below we offer several precautions that may be taken to minimize such artifacts. See our publications page for more information.