Automated Screening and Computer Assisted Diagnosis
Much of my recent research has been centered around segmentation, detection, and classification applications for medical images. Recent and ongoing projects include:
- Epithelial segmentation and classification of different types of odontogenic cysts in digital histopathology images.
- Segmentation of ovarian follicles in ultrasonographic images
- Automatic detection of polycystic ovaries in ultrasonographic images.
- Detection and segmentation of ovarian corpora lutea in ultrasonographic images.
Robust Analysis of Semi-automated Segmentation Algorithms
Working with difficult segmentation problems such as some of the problems above motivates the use of semiautomatic segmentation algorithms. We are particularly interested in algorithms where there is an interactive phase where an operator enters some high-level contextual information which is followed by a fully automatic phase. Most evaluations of such algorithms are performed by having a small number of human operators segment each image once, or just a few times. This limits the number of interactive inputs that are encountered by the algorithm, and therefore the confidence and accuracy of the evaluation. We are investigating methods of overcoming these difficulties. We are also interested in rigorously evaluating the relationships between segmentation accuracy, precision (reproducability), and operator time and the method by which contextual information is input (efficiency).
Digital Staining
Classical histologic techniques allow pathologists to visually distinguish different kinds of cells by chemically staining thin sections of biopsy tissue mounted on microscopic slides. It may be possible to obviate the need for a biopsy by performing an ultrasound scan and digitally “staining” the image with global filtering processes that exploit subtle differences in echotexture to make different kinds of tissue appear more different in ultrasound images. Indeed, recent work has shown that certain kinds of global filtering improve not only the difference between the visual appearance of the echotexture of luteal (glandular) tissue compared with stroma (vascular and connective tissue) in human ovaries but also the difference between statistical texture descriptors of said echotextures. This may achieved without knowledge of where each kind of tissue appears in the image (ie. without prior segmentation).