BIOE 484

Credit: 3 OR 4 hours.

Biomedical image data often come in extreme numbers: there is either so many of them that humans can't analyze them in reasonable time (e.g., three-dimensional light sheet microscopy data) or they are few, highly varied and of limited spatial and intensity resolutions (e.g., positron emission tomography scans). Furthermore, the extraction of image features and the characterization of modality-dependent background noise can be particularly challenging in typical biomedical scenarios. In this course, several applications of statistical learning to biomedical image data will be covered in depth from first principles. Analyses will be done in Python using the Scikit-learn package and all homework assignments comprise statistical analyses of biomedical image data in real decision scenarios. Histogram transforms and the fundamental properties of image texture will be introduced and revisited throughout the course. The extraction of both low- and high-order spatial features at multiple scales will be demonstrated and employed throughout the course. Support vector machines will be introduced and applied to image classification and interpretation tasks. The random forest algorithm will be introduced and used on a number of large- and small-data tasks. Multiple linear regression will be applied to neuroimaging data and some common methods of assessing model robustness shown. Cross-validation of image-derived decisions and some common methods of assessing model robustness will be shown. Feature selection and dimensionality will be discussed in terms of diagnostic task performance. The effects of inter-feature correlation upon prediction confidence will be discussed. Principal component analysis will be described and applied to various image processing tasks. Unsupervised clustering and cluster analysis of extracted image features will be introduced. Stochastic object models will be introduced and applied in various validation tasks.

3 undergraduate hours. 4 graduate hours. Prerequisite: BIOE 485 or permission of the instructor. Students are expected to be familiar with calculus, basic probability & sampling, vector spaces, matrix algebra and constrained optimization. Several NumPy objects and manipulations will be reviewed, and all necessary Sci-kit functions introduced; however, students are expected to have substantial experience with Python programming as the basics of such will not be covered.