Our goal is to create radiogenomics maps between quantitative image and genomic features for glioblastoma multiforme (GBM) and to assess the prognostic association of significant correlations.

To accomplish this we obtained multi-omics data from 251 patients and MR image data from a subset of 55 patients in the Cancer Genome Atlas (TCGA) and The Cancer Imaging Archive (TCIA) GBM databases.A board certified neuroradiologist traced 2D regions of interest (ROI) around necrotic and enhanced parts of the largest lesion in a selected slice from a T1 post-contrast MR, and around the region of hyperintensity obtained from the enhancement on the matched T2 FLAIR slice. These ROIs were used to compute quantitative image features from their shapes and pixel values.

We used a module network algorithm that integrates copy number, DNA methylation and gene expression data into 100 co-expressed gene modules. We established a radiogenomics map by correlating these modules with the quantitative image features, and used significant module-image feature correlation for survival analysis using Cox proportional hazards modeling.

A total of 28 quantitative image features were extracted for each of the necrosis, enhancement and edema ROIs in each patient. The radiogenomics map between modules and quantitative image features revealed 14, 10 and 16 significant gene-module associations with necrosis, enhancement and edema ROIs respectively.

For example we found a significant correlation between Module 64, enriched with genes in neuronal differentiation, and the compactness of the necrosis (p=0.0145). Also, we found that the amount of necrosis vs. enhancement or edema is correlated with Module 74, enriched in metabolism related genes (p<0.01). Finally, we found e.g. that the compactness of the necrosis ROI is correlated with poor survival (p=0.037).

Creating radiogenomics maps provides multi-scale insights by associating image features with molecular function. Moreover, these maps may provide additional insight for image features with prognostic correlations.

Modern multi-dimensional imaging in radiology yields much more information than the naked eye can appreciate. As a result, errors and variations in interpretations are currently the weakest aspect of clinical imaging. Computerized image analysis may provide solutions for ensuring the quality of medical image interpretation by yielding exhaustive, comprehensive and reproducible analysis of imaging features that are difficult to recognize with the naked eye. 

We developed computational models of multi-dimensional morphological properties of biomedical tissue. The Riesz transform and support vector machines are used to learn the organization of image scales and directions that is specific to a given biomedical tissue type. The models obtained can be “steered” analytically to enable rotation-covariant image analysis. While most rotation-invariant approaches discard precious information about image directions, rotation-covariant analysis enables modeling the local organization of image directions independently from their global orientation. Although already being important in 2D images, this becomes crucial to adequately leverage directional image information in 3D images. 

The proposed approach is first evaluated and compared to other state-of-the-art methods using benchmark collections of non-medical natural textures. Experimental evaluation reveals high classification accuracies for even orders of the Riesz transform, and suggests high robustness to changes in global image orientation and illumination. In a second step, we demonstrate the ability of the framework to model biomedical tissue. The scenarios investigated are (A) the prediction of descriptive terms about liver lesions seen in CT images and (B) the prediction of lung tumor recurrence from the texture appearance of components of lung nodules in CT images.

Overall, the proposed computational models were able to fit a wide range of textures and tissue structures. They can be trained to segment, classify and retrieve desired tissue categories. We are currently extending the framework to 3D, which is expected to fully leverage the wealth of modern radiology images.

In this talk I will describe our work over the years on high-content analysis of histology images. By high-content analysis, I allude to the extraction of structures and features while being cognizant of the organization of tissue, the microscopic context, and the underlying pathology. I will describe a few case studies that also include workflows for processing histology images. 

In one case study I will describe our integrative approach to associate morphological features with genomic data for use in prognostic biomarkers for a class of cancers. Our integrative approach will use new techniques of consensus clustering that allows the joint clustering of data from multiple modalities. 

In related studies, I will demonstrate how similar tools and workflows can be used to visualize cellular compartments, context and organization in the tissue microenvironment and also gross changes in specific tissue layers. The staple algorithms employed in these workflows achieve the registration and multi-level segmentation of given images. For each study, I will make the case for the novel workflows that were constructed to answer questions of biological or clinical importance. When appropriate, I will also provide details of salient algorithms realized in the workflows.

The healthcare environment is rapidly evolving, with increasing demand from payers, patients, and referring clinicians for radiologists to improve and demonstrate quality. However, quality in healthcare can be difficult to measure, let alone improve. 

Modern quality improvement methodologies have come to view quality improvement as synonymous with process improvement. Since radiology processes are based on information systems, to a significant degree, process improvement relies on improvement in electronic workflows. 

Furthermore, information systems allow for the efficient and rapid acquisition and display of performance data. Thus, there is a strong interplay between quality improvement and imaging informatics, bringing significant opportunities for the two to work together to shape the future of radiology.

Modern imaging techniques allows us to obtain high-resolution 2D and 3D images that are very rich in information. To be able to understand, compare different images, and automate analysis we need to find a subset of significant, repeatable features that summarize the data. Computer image analysis allows us to create unbiased and deterministic models that can be applied to these images to extract important features to characterize the studies.

We developed a 2D and 3D modular computerized Quantitative Image Feature extraction pipeline which extracts intensity, shape, edge, and texture features from a collection of DICOM images. These features then are used to summarize each case and further analyzed in other medical studies. While this is really helpful in 2D images, it is crucial in 3D volumes as visualizing 3D structures is a much harder problem.

To analyze the robustness of the algorithms used in our pipeline, we exposed the pipeline to the same set of images but with different segmentations. We observed that the sensitivity of most of the features to small segmentation errors remained low.

Our pipeline architecture is inherently modular, allowing more features to be added on the fly and therefore we are able to extend it as new features are developed.

The 2D pipeline has been applied successfully to liver and lung CT-images, the 3D pipeline is being used to analyze tumors in 3D lung CT volumes. An online version of the 2D pipeline exists now which allows users to submit their study cases and obtain a list of quantitative features. We are planning on releasing an online version of the 3D pipeline in the future, as well as integrating it with our in-house EPAD annotation system.

Glioblastoma is the most common and most aggressive primary human brain cancer in adults, with a 2-year survival rate of 20%. Heterogeneous phenotypes and clinical outcomes exist even within subtypes defined by molecular criteria. Magnetic Resonance Imaging (MRI) depicts the anatomic location of tumor with high resolution, and specific locations of tumor may associate with particular phenotypic features such as growth rate, invasiveness, and clinical outcome. Dynamic susceptibility contrast enhanced T2*-weighted MR perfusion characterizes hemodynamic aspect of the tumor, potentially providing a deeper biological understanding of glioblastoma. 

In this talk, I will discuss our work on extracting glioblastoma location features from T1-weighted MR and perfusion features from DSC T2*-weighted MR, and associating these quantitative features to patient prognosis. 

To study tumor location, we developed and applied automated computational image analysis to the MRIs of two large cohorts of glioblastoma. We identified a subgroup of tumors whose location associated with a distinct molecular profiles and clinical outcome. 

To study perfusion, we applied a novel analysis approach to the TCGA cohort, and correlated perfusion measurements with vascular and angiogenic gene expression. 

We also investigated the reproducibility of perfusion measurements generated by different software packages as well as ROIs delineated by different readers, and assessed the degree to which these variations impact clinical associations.

The ability to identify—early in the course of therapy—patients that are not responding to a given therapeutic regimen is highly significant. In addition to limiting patients’ exposure to the toxicities associated with unsuccessful therapies, it would allow patients the opportunity to switch to a potentially more efficacious treatment. 

In this presentation, we will discuss ongoing efforts at developing and applying advanced imaging technologies in order to predict, early in the course of therapy, the eventual response of individual patients to anti-cancer treatments. We will then describe how these data can be used to initialize and constrain predictive biophysical and biomathematical models of tumor growth and treatment response. 

We will conclude with a look towards the future on how this approach can enable the development of a field of “theoretical oncology”.

Computers are now extensively used throughout science, engineering, and medicine. Advances in computing allow researchers to build and test models of increasingly complex phenomena and thus to generate unprecedented amounts of data. These advances have created the need to make corresponding progress inour ability to understand large amounts of data and information arising from multiple sources. In fact, to effectively understand and make use of the vast amounts of information being produced is one of the greatest scientific challenges of the 21st Century.

Visual computing, which relies on and takes advantage of, the interplay among techniques of visualization, large-scale computing, data management, and imaging, is fundamental to understanding models of complex phenomena, which are often multi-disciplinary in nature. In this talk, I will provide examples of visual computing as applied to important problems in biomedicine.

Keynote Speaker for ISIS Annual Retreat

Identification of molecular subtypes has become critical in breast cancer management as many new therapies have been developed and proven effective for specific subtypes. As the molecular characterization of breast cancer continues to influence individual patient management, there is an urgent need to develop methods that integrate molecular and population-level data to evaluate the efficacy of current and emerging screening and treatment protocols to support the development of guidelines that reduce the burden of cancer in the general population.

In this talk, I will present my work on integrating molecular-markers on breast cancer progression models. In particular, we will discuss several approaches developed to estimate parameters that characterize progression and underlying survival by estrogen-receptor (ER) and human epidermal growth factor 2 (HER2). Furthermore, I will also show how these parameters are used in the development of molecular-specific models that are capable of evaluating the effects of screening, treatment and menopausal hormonal therapy (MHT) on breast cancer incidence and mortality.