1 / 68

Mixed-Integer Evolution Strategies and their Application to Medical Image Analysis

Mixed-Integer Evolution Strategies and their Application to Medical Image Analysis. Jeroen Eggermont Rui Li & Michael Emmerich. Background. Student Computer Science, Leiden University Master Thesis: “Rule-extraction and Learning in the BP-SOM Architecture” PhD Student, Leiden University

Download Presentation

Mixed-Integer Evolution Strategies and their Application to Medical Image Analysis

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Mixed-Integer Evolution Strategies and their Application to Medical Image Analysis Jeroen Eggermont Rui Li & Michael Emmerich

  2. Background • Student Computer Science, Leiden University • Master Thesis: “Rule-extraction and Learning in the BP-SOM Architecture” • PhD Student, Leiden University • PhD Thesis: “Data Mining using Genetic Programming: Classification and Symbolic Regression” • Postdoc, Division of Image Processing, LUMC • NWO-Savage Project Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  3. Overview • LKEB / Division of Image Processing • Atherosclerosis • NWO-Savage Project • Mixed-Integer Evolution Strategies • MIES Extensions • Fitness Based Partitioning • RBFN • Niching Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  4. LKEB / Division of Image Processing • Research entity of the Department of Radiology of the LUMC • The mission is directed towards the research, implementation and validation of automated segmentation, quantification and visualization approaches for structures in medical images • The objective results of these approaches serve to support the clinical decision making processes and the outcomes of clinical research studies Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  5. LKEB / Division of Image Processing • Core competence • Translation of medical imaging problems into clinically applicable technical solutions • Development of scientifically innovative automated analysis methods for medical images • Development of “physician friendly” user interfaces and visualization tools • Extensive clinical validation, beta testing and scientific publication Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  6. Research Development • Hospitals • Research institutes Application Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  7. LKEB: Image Modalities • Röntgen (X-ray) • Echocardiography • IntraVasculaire UltraSound (IVUS) • Magnetic Resonance Imaging (MRI) • Computer Tomography (CT) Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  8. LKEB: Topics • Cardiovascular (heart & bloodvessels) • Lungs (cpod, emphysema) • Orthopedics • Neurology (brains, Alzheimer disease) Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  9. Overview • LKEB • Atherosclerosis • NWO-Savage Project • Mixed-Integer Evolution Strategies • MIES Extensions • Fitness Based Partitioning • RBFN • Niching Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  10. Atherosclerosis Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  11. Atherosclerosis: CTA Analysis Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  12. Atherosclerosis: CTA Analysis Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  13. Lumen Calcified plaque CTA: Detecting Narrowings Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  14. Overview • LKEB • Atherosclerosis • NWO-Savage Project • Mixed-Integer Evolution Strategies • MIES Extensions • Fitness Based Partitioning • RBFN • Niching Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  15. ? Savage Project Feature Detector Images parameter setting Result So you set the first value to 0.1 the second to 42 … Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  16. Savage Project • How to find the optimal parameter setting ? • Trial and Error ? • Educated guess ? • Experience ?  How to optimize a new algorithm ? Solution: Optimize parameters automatically How ?  EvolutionaryAlgorithms Self-Adaptive Vision Agents using Genetic Evolution Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  17. Overview • LKEB • Atherosclerosis • NWO-Savage Project • Mixed-Integer Evolution Strategies • MIES Extensions • Fitness Based Partitioning • RBFN • Niching Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  18. Evolution Strategies: quick overview* • Developed: Germany in the 1970’s • Early names: I. Rechenberg, H.-P. Schwefel • Typically applied to: • numerical optimisation • Attributed features: • fast • good optimizer for real-valued optimisation • relatively much theory • Special: • self-adaptation of (mutation) parameters standard *A.E. Eiben and J.E. Smith, Introduction to Evolutionary Computing, Springer, 2003, ISBN 3-540-40184-9 Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  19. Evolution Strategies: technical summary* *A.E. Eiben and J.E. Smith, Introduction to Evolutionary Computing, Springer, 2003, ISBN 3-540-40184-9 Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  20. Mixed-Integer Evolution Strategies Why Mixed-Integer Evolution Strategies ?: Image segmentation algorithms have parameters of different types: • Ordinal Discrete {red, green, blue}, {true, false} • Continuous { 3.14, 6.28 } • Integer { 3,18, 42 }  Mixed-Integer Evolution Strategies have Self-Adaptive Mutation operators for each type  MIES outperformsstandard ES on Mixed-Integer Problems Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  21. Mixed-Integer Evolution Strategies Courtesy of Gunther Rudolph Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  22. Mixed-Integer Evolution Strategies CTA Lumen Detector Images parameter solution Expert contours optimal parameter solution Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  23. Mixed-Integer Evolution Strategies Fitness Function for each image I: • resample expert contour E and solution contour C every 2 degrees (180 points per contour) • compute average Euclidean distance between corresponding points Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  24. Mixed-Integer Evolution Strategies Experiments • 9 CTA data sets ( 59-82 images each) • 10 random seeds • (4+28)-MIES (4 parents, 28 offspring) • 50 generations • Comparison to the default parameter settings currently used Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  25. Mixed-Integer Evolution Strategies Conclusion:  MIES works and converges to solutions which are better than the default parameter settings. THERE IS NOT ONE OPTIMAL SOLUTION FOR ALL IMAGES ! Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  26. Overview • LKEB • Atherosclerosis • NWO-Savage Project • Mixed-Integer Evolution Strategies • MIES Extensions • Fitness Based Partitioning • RBFN • Niching Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  27. Fitness Based Partitioning Goal:We want to find optimal parameter settings for all images. • First Idea: Cluster images according to their image segmentation context • Problem: • We don’t know the number of image segmentation contexts • We have no natural distance measure that indicates which images require similar parameter solutions • Second Idea: Co-evolve parameter solutions and image clusters • Problem: • This requires a large number of fitness evaluations  Image segmentation is very time consuming Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  28. Fitness Based Partitioning Group images together that require similar parameter settings for image segmentation Find optimal parameter settings for these groups Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  29. ? ? ? Fitness Based Partitioning Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  30. assign images to partitions CTA Lumen Detector CTA Lumen Detector MI-ES 1 MI-ES 2 best solution best solution CTA Lumen Detector CTA Lumen Detector Fitness Based Partitioning Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  31. Fitness Based Partitioning: Artificial Problems Uniform distribution Gaussian distribution Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  32. Fitness Based Partitioning: Test Results • T = number of MI-ES generations. • Types of Errors: - a sample put on wrong island, or - combined & split partitions Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  33. Fitness Based Partitioning: Potential Problem Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  34. Fitness Based Partitioning: Potential Problem 2 Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  35. Fitness Based Partitioning: Empty Partitions Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  36. Fitness Based Partitioning What to when a partition becomes empty ? • destroy MI-ES population • clone population of the MI-ES algorithm of the largest partition • divide images of largest partition CTA Lumen Detector CTA Lumen Detector MI-ES 1 MI-ES 2 MI-ES 1’ best solution best solution Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  37. Fitness Based Partitioning • Divide images over K partitions • Initialize a MIES algorithm for each partition • for T iterations do • Apply a MIES to each partition for G generations • Select the best MIES solution for each partition • Test the best MIES solution of each partition on all problem instances • Re-assign images to the partition with the best solution • Split partitions if needed Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  38. Fitness Based Partitioning Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  39. Fitness Based Partitioning: Experiments • 9 CTA Data sets • 10 random seeds • (4+28)-MIES algorithm (4 parents, 28 offspring) • each MIES algorithm is run for 5 generations at a time • images are reassigned 10 times Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  40. Fitness Based Partitioning: Results 2-partitions expert contour evolved solution contour best solution MIES 1 best solution MIES 2 Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  41. Fitness Based Partitioning: Results Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  42. Fitness Based Partitioning: Results overall fitness results Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  43. Fitness Based Partitioning: Conclusions • F.B.P. results in parameter settings which lead to better CTA lumen segmentations • Performance assessment: • High reliability: Groups of images usually assigned to the same partition • ….however there remains some sensitivity to the random seed used. Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  44. Overview • LKEB • Atherosclerosis • NWO-Savage Project • Mixed-Integer Evolution Strategies • MIES Extensions • Fitness Based Partitioning • RBFN • Niching Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  45. MIES+RBFN: Time expensive evaluations CTA Lumen Detector Images parameter solution Expert contours optimal parameter solution Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  46. MIES+RBFN • What if we could predict quality of a parameter solution • How ? • Neural Networks (RBFN) • Image processing is time consuming • Every potential parameter solutions must be evaluated Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  47. MIES+RBFN: Train NN CTA Lumen Detector Images parameter solution Expert contours optimal parameter solution Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  48. MIES+RBFN CTA Lumen Detector Images parameter solution optimal parameter solution Expert contours Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  49. MIES+RBFN Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

  50. MIES+RBFN Mixed-Integer Evolution Strategies and their application to Medical Image Analysis

More Related