An approach to analyzing histology segmentations using shape distributions

A master's thesis by Jasper Zhang
submitted to the faculty of Drexel University
in partial fulfillment of the requirements for
the degree of Masters of Science
in Computer Science
March 2008


Abstract

Histological images are the key ingredients in medical diagnosis and prognosis in today's medical field. They are imagery acquired by analysts from microscopy to determine the cellular structure and composition of a patient's biopsy. This thesis provides an approach to analyze the histological segmentation obtained from histological images using shape distributions and provides a computationally feasible method to predict their histological grade.

This process provides a way of generating suggestions using segmented images in a way that is independent of the segmentation process. The process generates histograms for each image that describes a set of shape distributions generated from eight metrics that we have devised. The shape distributions are extracted from a learning set that the user provides. The shape distributions are then analyzed by querying a classification for each case using K-nearest-neighbor. The quality of the classifications is measured by a composite measure composed of precision and recall based on the query.

Thesis

PDF: download
Word (*.doc): download

Raw Images

Grade 1-3 cases 1-5: download
Grade 1-3 cases 6-10: download (to come)

Segmented Images

Complete set: download

Histograms

Complete pre-filter set: download
Complete post-filter set: download

Sliding Window Analysis

(to come)

Implementation details

Shape functions

(to come)

Utilities

(to come)

Analysis

(to come)


Jasper Zhang (jzz22 _AT_ drexel.edu), 2008