Schließen der semantischen Lücke in der medizinischen Bildanalyse : kontextabhängige Objektextraktion aus hierarchisch partitionierten Bildern

In any biomedical domain there are applications in which objects are identified, measured and counted. Examples are the identification and staging of tumors or counting of cells in micrographs for cytological screenings. In the long run the constantly increasing number of acquired images in biomedical applications will only be evaluable by computerized aid. In contrast to the claims of medical image processing such a computerized object extraction is not fully integrated in clinical routine. This work examines the reasons for this observations and proposes an integrated framework as a pragmatic solution. Its application principle is demonstrated by use-cases. Object extraction is based on formal representation of user knowledge by means of image processing. This requires technical knowledge of methods on the one hand while on the other hand contextual knowledge of the application is needed. For this task no general solution has been found within the last 40 years of research in image analysis. Moreover there exist many sophisticated segmentation algorithms and paradigms for object modeling. Presumably there is no such general solution since the semantic gap opens between formal and contextual representation of knowledge. However the increasing number of images requires automated approaches and there exist many application specific solutions modeling knowledge on different levels of abstraction. But at any time the imaging modality or object under investigation is modified, methods need to be re-parameterized and even re-implemented by an image processing expert. The challenge lies within the heuristics of knowledge representation itself which is based on the trade-off between accuracy, generalization and the effort needed for parameterization and evaluation of results. The framework proposed in this work provides a conceptual separation of low-level image processing and high-level object description. At the same time the expression of knowledge of the user and the developer is considered by appropriate views. Low-level image processing is done by multi scale image decomposition into all visually plausible image regions. These are organized in a tree structure and for each region a set of descriptive features such as the mean grey value or the Fourier descriptors of its contour is computed. Thus a multidimensional feature space of region descriptions is formed. Within this new approach high-level object description becomes a classification task to be solved by classical means of pattern recognition and machine learning. Here application specific knowledge is provided by training data and selection and parameterization of the respective classifier. In addition, the framework fully integrates result evaluation in the development and application view. All components of the framework are exchangeable in the development view and thus new methods are designed and integrated without re-implementation of the flow of operations, interfaces and already developed object descriptions. For both multi scale decomposition and classification there are several operations available. Additionally a domain specific language is provided for feature description and a virtual machine is integrated for online computation. Thus no recompilation of the software is required during application. Specific descriptions of objects are learned in the application view. Here a trained method is applied to a given image repository and its extraction result is rated without deeper knowledge of the actual method and its implementation. The level of actual abstraction of the learning algorithm remains the uncloseable opening of the semantic gap. Thus a minimum of technical knowledge will always be required. For this purpose a new interval classifier on the multi scale decomposition is used besides well established classifiers from the literature. The application of the framework is demonstrated by five sample use-cases: At first four cases demonstrate the systematic organization of operations by extraction of metacarpal bones from clinical hand radiographs. Parameterization of multi scale decomposition, feature extraction, training of classifiers and an application scenario are tested. In the fifth use-case the integration of the framework into two complex applications of image analysis is demonstrated. Firstly cells are identified in fluoroscopic micrographs and then extraction of tongue regions from a series of medial sagittal NMR images of the head is performed. In this work it is demonstrated that the framework provides a tool to bridge the semantic gap by mapping contextual knowledge of the user onto complex image processing with a minimum of technical knowledge. It becomes automatically and reproducible applicable to series of images from various modalities and applications. Even although this approach is heuristic it yields a closed and systematic approach to an urgent problem.

Grantor dc:publisher

Publikationsserver der RWTH Aachen University

Year dc:date

2007

Author dc:creator

• Thies, Christian

Contributors dc:contributor

• Deserno, Thomas

• info:eu-repo/classification/ddc/004
• Wissensrepräsentation
• Bildanalyse
• Objekterkennung
• Data Mining
• Informatik
• Multiskalenanalyse
• inhaltsbasierte Bildsuche
• biomedizinische Objekte
• Object recognition
• image analysis
• content based image retrieval
• multiscale analysis

Statement dc:rights

• info:eu-repo/semantics/openAccess

Language dc:language

ger

Identifier

https://publications.rwth-aachen.de/search?p=id:%22RWTH-CONV-124060%22

OAI identifier oai:identifier

oai:publications.rwth-aachen.de:62494