Medical image analysis has a long history, and recent years have included several groundbreaking studies in histopathology. Below is a broad overview of the steps required to quantify and analyze medical imaging, along with our group’s work in developing novel methods and experimental models.
Multi-Scale Framework Development
We often deal with very high-resolution images that contain large amounts of variance and noise. We are interested in developing frameworks that can deal with whole-slide images on the order of 5-10 GB in size, identifying important information at a variety of scales (sample, tissue, region, cell, and intra-cell) using fast and robust methods, and standardizing images with respect to lighting, stain intensity, and noise.
Finding Biologically Important Structures
Object segmentation is a major topic in image analysis. Segmentation can be done in a low-level manner driven by image pixels and homogeneity constraints, or at a higher level by integrating domain knowledge to identify relevant areas and eliminate unnecessary ones. In biomedical imaging, both approaches have their uses in identifying important biological constructs such as tissue regions, glands, cells, and sub-cellular components.
Quantifying Biological Domain Knowledge
Feature design and extraction is a complex process that involves two phases: First, we must identify the (qualitative) characteristics of the images and segmented structures that are necessary to identify the target classes, which could be disease presence, cancer severity, patient outcome, etc. Second, we need to ensure that we design feature operators that reliably and robustly quantify those characteristics. Features should be independent from one another (that is, each feature should add new information to the dataset) and should capture intuitive biological processes.
Viewing and Operating on Quantitative Data
Once the features are extracted for each object of interest (image, patient, nucleus, etc.) we can analyze them prior to building a fully-fledged predictive model. For example, we can identify feature groups that are highly correlated, indicating that we might be extracting redundant information from the system. We can perform “feature selection”, which determines the feature’s ability to encode information about the target classes – this serves as a nice sanity check that our feature extraction is in fact yielding useful information. Finally, we can look at the feature values themselves; this topic is covered in great detail in our visualization projects.
All Categories