46. ANALYZING PATHOLOGY IMAGES

Department: Computer Science & Engineering
Faculty Advisor(s): Yoav Freund

Primary Student
Name: Mayank Mahesh Kabra
Email: mkabra@ucsd.edu
Phone: 858-534-8822
Grad Year: 2010

Abstract
Development of high resolution whole slide microscopic digital scanner marks a shift in histopathology, as pathologists move away from microscopes towards computer screens. With this shift, image processing techniques can now be used to speed up the diagnosis. But an unnoticed aspect is the ability to store images digitally instead of physically and the ability to access them remotely reducing the tedium of collaboration and second opinions encountered with physical slides.

But size of the digital images is a serious impediment in achieving the full potential of digital histopathology. A single image scanned at 40x resolution requires 1GB (after jpeg compression). A terabyte hard disk can thus store only a thousand or so images. While sharing it over the network may take quite a substantial time making the whole system less interactive and less useful.

Pathologists, however, hardly ever view the whole image at high resolution but concentrate on few relevant parts. Relevant parts, if stored at higher resolution can save space, and if transmitted earlier over network can reduce delay.

We chose the task of analyzing prostate images as its cancer is most widespread. In prostate cancer, glandular cells located on the periphery of glands, divide at faster rate and also break away from the periphery of glands destroying the gland structure. Presence of unusually high number of nuclei indicate faster dividing regions while destruction of glandular structure is partially indicated by presence of smaller glands. We use machine learning techniques to learn these features to detect cancer.

« Back to Posters or Search Results