An Analysis of Scale and Rotation Invariance in the Bag-of-Features Method for Histopathological Image Classification

• Type of Material: Article
• Authors: S. Hussain Raza, Georgia Institute of Technology; R. Mitchell Parry, Emory University; Richard A. Moffitt, Emory University; Andrew Young, Emory University; Dongmei Wang, Emory University
• Language: English
• Date: 2011-01-01
• Publisher: SPRINGER-VERLAG BERLIN
• Publication Version: Author Accepted Manuscript (After Peer Review)
• Copyright Statement: Med Image Comput Comput Assist Interv
• Final Published Version (URL): http://dx.doi.org/10.1007/978-3-642-23626-6_9
• Title of Journal or Parent Work: MEDICAL IMAGE COMPUTING AND COMPUTER-ASSISTED INTERVENTION, MICCAI 2011, PT III
• Volume: 6893
• Issue: Pt 3
• Start Page: 66
• End Page: +
• Grant/Funding Information: This research has been supported by grants from National Institutes of Health (Bioengineering Research Partnership R01CA108468, P20GM072069, Center for Cancer Nanotechnology Excellence U54CA119338), Georgia Cancer Coalition (Distinguished Cancer Scholar Award to Professor Wang), Hewlett Packard, and Microsoft Research.
• Keywords: Life Sciences & Biomedicine; Texton-based Approach; Radiology, Nuclear Medicine & Medical Imaging; Computer Science, Theory & Methods; Computer Aided Diagnosis; Bag-of-features Method; DIAGNOSIS; Technology; Image Classification; Computer Science; Science & Technology
• Research Categories: Health Sciences, Radiology

Abstract

The bag-of-features method has emerged as a useful and flexible tool that can capture medically relevant image characteristics. In this paper, we study the effect of scale and rotation invariance in the bag-of-features framework for Renal Cell Carcinoma subtype classification. We estimated the performance of different features by linear support vector machine over 10 iterations of 3-fold cross validation. For a very heterogeneous dataset labeled by an expert pathologist, we achieve a classification accuracy of 88% with four subtypes. Our study shows that rotation invariance is more important than scale invariance but combining both properties gives better classification performance.