Unsupervised medical image classification by combining case-based classifiers Thien Anh Dinh 1 , Tomi Silander 1 , Bolan Su 1 , Tianxia Gong Boon Chuan Pang 2 , Tchoyoson Lim 2 , Cheng Kiang Lee 2 Chew Lim Tan 1 ,Tze-Yun Leong 1 1 National University of Singapore 2 National Neuroscience Institute 3 Bioinformatics Institute, Singapore
19
Embed
Unsupervised medical image classification by combining case -based classifiers
Thien Anh Dinh 1 , Tomi Silander 1 , Bolan Su 1 , Tianxia Gong Boon Chuan Pang 2 , Tchoyoson Lim 2 , Cheng Kiang Lee 2 Chew Lim Tan 1 ,Tze-Yun Leong 1 1 National University of Singapore 2 National Neuroscience Institute 3 Bioinformatics Institute, Singapore. - PowerPoint PPT Presentation
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Unsupervised medical image classification by combining
case-based classifiersThien Anh Dinh1, Tomi Silander1, Bolan Su1, Tianxia GongBoon Chuan Pang2, Tchoyoson Lim2, Cheng Kiang Lee2
Chew Lim Tan1,Tze-Yun Leong1
1National University of Singapore2National Neuroscience Institute
3Bioinformatics Institute, Singapore
2
Automated medical image annotation
• Huge amount of valuable data available in medical image databases
• Not fully utilized for medical treatment, research and education
• Medical image annotation:1. To extract knowledge from images to facilitate text-
based retrieval of relevant images2. To provide a second source of opinions for clinicians
on abnormality detection and pathology classification
3
Problem
• Flowchart of current methods
• Challenges in current methods• Highly sensitive and accurate segmentation• Extracting domain knowledge• Automatic feature selection
• Time-consuming manual adjustment process reduces usages of medical image annotation systems
Extracting features
Selecting discriminative features
Building classifiers Labeling
4
Objective
• An automated pathology classification system for volumetric brain image slices
• Main highlights
1. Eliminates the need for segmentation and semantic or annotation-based feature selection
• Reduces the amount of manual work for constructing an annotation system
2. Extracts automatically and efficiently knowledge from images
3. Improves the utilization of medical image databases
5
System overview• Case-based classifier
• Gabor filters • Non domain specific
features• Localized low-level
features
• Ensemble learning• Set of classifiers• Each classifier with a
random subset of features• Final classification: an
aggregated result
6
Sparse representation-based classifier
• Sparse representation-based classifier (SRC) proposed by Wright et al. for face recognition task
• Non-parametric sparse representation classifier
• SRC consists of two stages1. Reconstructing: a test image as a linear
combination of a small number of training images2. Classifying: evaluating how the images
belonging to different classes contribute to the reconstruction of the test image
7
Sparse representation-based classifier(Wright et al.)
r1 = || y – (a7x7 + a172x172 + a132x134)||2r2 = || y – (a23x23 + a903x903)||2
9
Ensemble of weak classifiers
• Combine multiple weak classifiers
• Take class specific residuals as confidence measuresThe smaller the residual for the class, the better we construct the test by just using the samples from that class
• To classify image y, compute average class-specific residuals of all W weak classifiers
10
Domain
• Automatically annotate CT brain images for traumatic brain injury (TBI)TBI: major cause of death and disability