289 - PowerPoint

Document Sample
289 - PowerPoint Powered By Docstoc
					Computational Image
   Classification
  UMBC Department of Computer Science
        eBiquity Research Group
           February 19, 2010
Overview
   Introductions
   Image Classification
   Initial Results
   Future Efforts
Introductions
   Faculty
       Yelena Yesha, PhD
       Michael Grasso, MD, PhD
       John Dorband, PhD
       Tim Finin, PhD
       Milt Halem, PhD
       Anupam Joshi, PhD
   Graduate students
       Ronil Mokashi
       Darshana Dalvi
Computational Image Classification
   Categorize a raster image into a finite
    set of classes.
       Convert raster data into feature vectors.
       Support vector machine image classifier.
       Metadata to map specific classes to
        biological characteristics.
Image Classification Examples
   Computer assisted diagnosis of prostate
    and breast cancer biopsies.
   Segmentation of hysteroscopy video.
   Echocardiogram analysis.
   Skin cancer detection.

           Biomedical Imaging: from Nano to Macro, 2007;:1284-1287
                                 IEEE TITB, 2008 May;12(3):366-376
                       Proceedings 27th IEEE EMBS, 2005;:5680-5683
                     Conf Proc IEEE Eng Med Biol Soc, 2006;1:4775-8
Related Efforts: Video Segmentation
   Laparoscopic cholecystectomy videos.
       378 representative images from 5 videos.
       Analyzed 49 separate image features.
Related Efforts: Video Segmentation
   Image classification.
        Distance metric to identify best features.
        Support vector machine image classifier.
        Accuracy of 91%.

        Video


        Segments

        Frames
Related Efforts: Video Segmentation
   Future directions.
       Real-time analysis to assess patient safety.
       Time and motion analysis of surgical
        instruments.
       Classification of pathology.
            Hiatal hernias.
            Capsule endoscopy.
Related Efforts: Cancer Screening
   Skin caner screening.
       Handheld iPhone image classifier.
       Tool for primary care physicians.
       Identify lesions in
        need of dermatology
        referral.
       NIH and UMB
        proposals pending.
    Image Classification Approach

                              Feature Extraction
                                                          Feature Models




Images Organized by Class                          Model Features




                               Feature Extraction
                                                          Image Classifier
              Unknown Image
Image Features
   Spectral features (color/tone).
       Histogram (3D, color, binary, gray).
       Distribution, size, width, mean, stdev.
       Do not vary with translation and rotation.
   Textural features (spatial distribution).
       Gray-level co-occurrence matrix (GLCM).
       Energy, entropy, contrast, correlation.
       Independent of color distribution.
IEEE Transaction on Systems, Man, and Cybernetics. 1973 Nov; 3(6):610-621
Advanced Features - Context
   Image segmentation.
       Regions of interest (contextual features).
       Threshold algorithms - Maximum Entropy,
        Otsu Threshold, Watershed, etc.
   Segmentation features of actin fibers.
       Density - Area, Mean Gray Level.
       Distribution - Centroid, Center of Mass.
       Orientation - Angle, Elliptical Fit (wrt cell).
       Order - Angle, Elliptical Fit (wrt fibers).
Advanced Features - Clustering
   K-means clustering of image features.
       Partitions images into clusters based on the
        nearest mean, based on a first-order Markov
        property.
       Based on the assumption that images with
        similar clinical features are more likely to be
        found in the same cluster.
Model Development
   Distance metrics
       Manhattan distance, Jeffrey divergence.
       Classification threshold.
   Support vector machines
       Machine learning methods.
       An N-dimensional hyperplane optimally
        separates images into categories.
       This mapping is performed by a set of
        mathematical functions, known as kernels.
Initial Results - Feature Analysis
   Initial image classification experiment.
       Evaluated 15 spectral and textural features.
       Total of 11 images in 4 groups.
            Focal adhesions images, actin stained.
            1hdry, 1hwet, 24hdry, 24hwet.
   Analysis.
       Leave-one-out technique over all 11 images.
       Manhattan distance.
       Threshold of 5% (0.3% for histograms).
Initial Results - Feature Analysis
           Trait+   Trait-            Promising features.
                                          Gray-scale distribution
Feature+    43       17      60
                                          Medium
Feature-    13       147     160          Mode
                                          Homogeneity
            56       164     220
                                          Energy
                                          Entropy
  Sensitivity = 76.8%                     Inverse difference
  Specificity = 89.6%                      moment
  Accuracy = 86.4%
Initial Results - Segmentation
   Second image classification experiment.
       Evaluating 4 new image features.
            Density, Distribution, Orientation, Order
       Experimenting with threshold algorithms to
        optimize image segmentation.
       Total of 11 images in 4 groups.
            Focal adhesions images, actin stained.
            1hdry, 1hwet, 24hdry, 24hwet.
     Initial Results - Segmentation




   Orientation feature using elliptical fit.
       Image moment-preserving threshold.
       Elliptical fit.
       Cell angle = 132°
       Average (weighted) actin fiber angle = 129°.
Initial Results - Segmentation
Initial Results - Analysis Platform
   To automate and optimize image
    processing algorithms.
Future Efforts
   Use feature analysis to develop a
    support vector machine image classifier.
   Continue image segmentation work.
   Correlate actin data to other images.
   Incorporate successful algorithms in the
    Analysis Platform.
   Identify ontologies to map specific
    classes to biological characteristics.

				
DOCUMENT INFO
Shared By:
Categories:
Stats:
views:21
posted:5/4/2011
language:English
pages:21