Machine Learning for Biomedical Image Analysis

Wednesday, April 14, 08:30 - 12:15

Presented by

Presented by Marco Loog (Delft University of Technology, The Netherlands) and David Tax (Delft University of Technology, The Netherlands)

Description

Machine learning and pattern recognition techniques utilize exemplary data rather than models to solve certain decision tasks. They enable their users to exploit various sources of data (images, annotations, general patient data, etc.) even if prior knowledge on the nature of these data and their interrelationship is limited or weak. More specifically, supervised machine learning and pattern recognition methods aim to "learn" a potentially complex input / output relation of interest from given example input / output pairs.

Machine learning plays a role of increasing importance in many biomedical image analysis and processing challenges. This tutorial offers a brief, and hopefully nonstandard, introduction to the basics of machine learning together with an overview of the key components and main considerations that play a role in good machine learning practice. The focus in this is on classification problems. While the basics are the topic of the first block, the second takes these as a prerequisite and presents various extensions to this standard, and in a sense limited, view of machine learning. It provides an overview of several techniques that have received increased attention in the past years.

Various concepts are illustrated by means of research examples and biomedical applications.

Part I: Established and Standard Techniques

A condensed overview of the standard classification pipeline is given. Various classifiers are touched upon (think linear discriminants, support vector machines, boosting, classifier combinations, ensemble methods, nearest neighbor methods), and representation and evaluation issues are briefly considered.

• The machine learning trinity
  • Representation
  • Generalization
  • Evaluation
• Bayes risk and density-based classifiers
  • LDA
  • QDA
• Distance based classifiers
  • kNN
  • NMC
• Similarity and dissimilarity approaches
  • SVM
  • kernels
  • dissimilarities
  • prototypes
  • structural representations
• Issues of complexity
  • bias / variance
  • learning curves
• The next hot thing?

Part II: Emerging and Less Standard Approaches

Deviations from the standard setting (from Part I) are considered and more realistic and, potentially, more powerful methods are discussed. We make a case for research into these "emerging" and less standard directions. Topics are multiple instance learning, semi-supervised learning, active learning, and learning under non-standard cost.

• One-class classification
• Multiple instance learning
• Semi-supervised learning
• Active learning
• Contextual classification
• Optimization of other risks

A Rather Biased List of References

• Arzhaeva, Tax, van Ginneken, “Dissimilarity-based classification in the absence of local ground truth: application to the diagnostic interpretation of chest radiographs”, Patt. Recog., 2009
• Arzhaeva, Tax, van Ginneken, “Improving computer-aided diagnosis of interstitial disease in chest radiographs by combining one-class and two-class classifiers”, SPIE, 2006
• Bishop, “Pattern Recognition and Machine Learning”, Springer, 2006
• Chapelle, Schölkopf, Zien, “Semi-Supervised Learning”, MIT Press, 2006
• Collier, “Priestess of Delphi”, 1891
• Cortes, Vapnik, “Support vector networks”, Mach. Learn., 1995
• Duda, Hart, Stork, “Pattern Classification”, John Willey & Sons, 2000
• Freund, “Boosting a weak learning algorithm by majority”, Information and computation, 1995
• van Ginneken, et al. “Comparing and combining algorithms for computer-aided detection of pulmonary nodules in computed tomography scans: the ANODE09 study”, Tech. Rep., 2010
• van Ginneken, Stegmann, Loog, “Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database”, MedIA, 2006
• van der Heijden, Duin, et al., “Classification, parameter estimation, and state estimation”, Wiley, 2004
• Juszczak, Duin, “Selective sampling based on the variation in label assignments”, ICPR, 2004
• Klein, Loog, et al., “Early diagnosis of dementia based on intersubject whole-brain dissimilarities”, ISBI, 2010
• Li, “Markov Random Field Modeling in Image Analysis”, Springer, 2009
• Loog, van Ginneken, “Supervised segmentation by iterated contextual pixel classification”, ICPR, 2002
• Loog, van Ginneken, “Static posterior probability fusion for signal detection”, ICPR, 2004
• Loog, van Ginneken, Duin, “Dimensionality reduction of image features using the canonical contextual correlation projection”, Patt. Recog., 2005
• Loog, van Ginneken, “Segmentation of the posterior ribs in chest radiographs using iterated contextual pixel classification”, TMI, 2006
• Loog, de Bruijne, “Discriminative shape alignment”, IPMI, 2009
• Maron, Lozano-Pérez, “A framework for multiple-instance learning”, NIPS, 1998
• Niemeijer, et al., “Comparative study of retinal vessel segmentation methods on a new publicly available database”, SPIE, 2004
• Pekalska, Duin, “The Dissimilarity Representation for Pattern Recognition”, World Scientific, 2005
• Raundahl, Loog, et al. “Automated effect-specific breast density measures”, TMI, 2008
• Sánchez, Niemeijer, et al., “Active learning approach for detection of hard exudates, cotton wool spots and drusen in retinal images”, SPIE, 2009
• Sánchez, Niemeijer, et al., “Improving hard exudate detection in retinal images through a combination of local and contextual information” , ISBI, 2010
• Schölkopf, Smola, “Learning with kernels”, MIT Press, 2002
• Settles “Active Learning Literature Survey”, Tech. Rep. 1648, 2009
• Sindhwani et al., “Beyond the point cloud: from transductive to semi-supervised learning”, ICML, 2005
• Smal, Loog, et al., “Quantitative comparison of spot detection methods in fluorescence microscopy”, TMI, 2010
• Sørensen, Shaker, de Bruijne, “Quantitative analysis of pulmonary emphysema using local binary patterns”, TMI, 2010
• Tax, Duin, “Support vector data description”, Mach. Learn., 2004
• Tax, Duin, “Learning curves for the analysis of multiple instance classifiers”, S+SSPR, 2008
• Tax, Loog, Duin, “Optimal mean-precision classifier”, MCS, 2009
• Tax, Duin, Arzhaeva, “Linear model combining by optimizing the Area under the ROC curve”, ICPR, 2006

Speaker Biographies

David M.J. Tax received his Ph.D. in 2001 at the Delft University of Technology in the Pattern Recognition Group, under the supervision of R.P.W. Duin and based on his thesis 'One-Class Classification'. Currently, he is an assistent professor in the Pattern Recognition Laboratory at Delft University of Technology. His main research interests are the learning and development of detection algorithms, the representation of data, simple and elegant classifiers and the fair evaluation of such methods.

Marco Loog received a Ph.D. degree from the Image Sciences Institute, NL, for the development and improvement of statistical pattern recognition methods and their use in the processing and analysis of images. After this joyful event, he moved to Copenhagen, DK, were he acted as assistant and, eventually, associate professor. After several splendid years, Marco moved to Delft University of Technology where he now works as an assistant professor in the Pattern Recognition Laboratory. Currently, his ever-evolving research interests include multiscale image analysis, semi-supervised and multi-instance learning, saliency, the dissimilarity approach, and black math.

©2011 Conference Management Services, Inc. -||- email: webmaster2010@biomedicalimaging.org -||- Last updated Thursday, April 08, 2010