DaUliKlank

Chair for Computer Aided Medical Procedures & Augmented Reality
Lehrstuhl für Informatikanwendungen in der Medizin & Augmented Reality

Diploma thesis

[Subtitle] Genetic Algorithms in Vision
Thesis by: Ulrich Klank
Advisor: Nassir Navab
Supervision by: Nicolas Padoy
Due date: 15th of May 2007

Abstract

This work addresses the recognition of phases in minimally invasive surgeries. These phases are sequential in time and should be recognized using observations which are acquired automatically. The endoscopic view of the surgeon provides information about the surgical field. This work focuses on the recognition of surgical phases using image features computed from these endoscopic images. It is a complex task to select efficient features and few literature exists about features that discriminate surgical phases in endoscopic views. In this work, a new Genetic Programming approach is proposed to automate the search for efficient features. A feature is modeled as a program and those programs are evolved to improve the recognition rate. For the representation of the programs a programming language was defined, specialized on computation of image features. Programs of this programming language are evaluated by executing them on a virtual machine with labeled sample images as input. Once the programs are evaluated and achieved a so called fitness, the best programs are selected as parents for a slightly changed and probably improved new generation of programs. Finally, the resulting features are compared with several standard image features, to show their performance in distinguishing between two phases using an image.With a selection of the best features, a multi-class classifier is built. It is compared with an early approach which is based on a neural network fed with a set of standard image features. entia non sunt multiplicanda praeter necessitatem (entities should not be multiplied beyond necessity) – William of Ockham, 14th century (Ockham’s razor) vii

Resources

Literature


Students.ProjectForm
Title: Automatic Feature Computation for Endoscopic Image Classification
Abstract: This work addresses the recognition of phases in minimally invasive surgeries. These phases are sequential in time and should be recognized using observations which are acquired automatically. The endoscopic view of the surgeon provides information about the surgical field. This work focuses on the recognition of surgical phases using image features computed from these endoscopic images. It is a complex task to select efficient features and few literature exists about features that discriminate surgical phases in endoscopic views. In this work, a new Genetic Programming approach is proposed to automate the search for efficient features. A feature is modeled as a program and those programs are evolved to improve the recognition rate. For the representation of the programs a programming language was defined, specialized on computation of image features. Programs of this programming language are evaluated by executing them on a virtual machine with labeled sample images as input. Once the programs are evaluated and achieved a so called fitness the best programs are selected as parents for a slightly changed and probably improved new generation of programs. Finally, the resulting features are compared with several standard image features to show their performance in distinguishing between two phases using an image.With a selection of the best features a multi-class classifier is built. It is compared with an early approach which is based on a neural network fed with a set of standard image features.
Student: Uli Klank
Director: Nassir Navab
Supervisor: Nicolas Padoy
Type: Diploma Thesis
Area: Surgical Workflow
Status: finished
Start: 2006/11/15
Finish: 2007/05/14


Edit | Attach | Refresh | Diffs | More | Revision r1.9 - 25 Sep 2007 - 17:22 - NicolasPadoy