Computational Framework for Prostate Cancer Diagnosis and Prognosis
Award year: 2006-2007
The paradigm of cancer detection and diagnosis currently consists of screening the population, followed by a biopsy to diagnose disease. Consider, for example, prostate cancer -- the most common internal cancer in U.S. males. While new diagnoses number ~232,000 annually (2005 estimate), biopsies for pathologic diagnoses are conducted on almost five times the number due to imperfect screening, repeat procedures and therapy follow-up. An estimated 20 million specimens place increased demands on pathology practice, provide opportunities for malpractice lawsuits and produce vexing research questions. Pathologic examinations consist exclusively of human examination, using light microscopy, of tissue stained with dyes -- a procedure that has remained relatively constant for decades. We have developed a new approach to microscopy in the mid-infrared (mid-IR) spectral region (2-12 ƒm wavelengths) using novel array detectors. In combining imaging with spectroscopic discrimination, we conduct Fourier transform infrared (FTIR) spectroscopic imaging in the manner of optical microscopy. As opposed to optical microscopy, however, no stains are required to achieve contrast in images of human tissue. Instead, spectral features allow for the imaging of specific molecular signals and provide contrast based on the chemical composition of the tissue. The data also provides a unique opportunity for automated cell types and disease recognition in complex tissue using learning algorithms.
With typical data sets in the gigabyte range, data retrieval, classification and visualization processes represent an enormous challenge and a critical barrier to using FTIR imaging data for computer-aided diagnoses (CADx). The exciting possibility of diagnosing cancer in an automated, objective and reproducible manner and without human intervention will not be realized unless these data management and computational challenges are overcome. Automated learning methods, as those developed in the Automated Learning Group at NCSA, provide a platform for addressing most aspects of CADx. Collaborating with this group and using NCSA resources, our specific goals are to: 1. Determine significant spectral features that enable automated diagnosis of malignancy in prostate biopsies 2. Identify spectral "signatures" of disease using patterns of spectral features and organize patients into self-similar groups based on these signatures 3. Test the significance of predicting outcome in patients based on spectral data and combinations with spectral data and clinical data The results will be useful in translational activities as well as providing clues to biological activity in various types of tumors. The integration of diverse information is a unique feature of the anticipated protocol and likely to be especially valuable. While the present project relates to prostate cancer, data analysis strategies developed in collaboration with NCSA will play a vital role in future applications of spectroscopic imaging to study biological transformations, both in our research and that of other groups. Last, the integration of clinical and spectral information with multimodal imaging will provide a novel cyberenvironment for medical diagnostics.