Classification

During our research years at the Pattern Recognition Lab of the Friedrich-Alexander University Erlangen-Nuremberg we have acquired a profound knowledge and experience in pattern extraction and classification.

In collaboration with the Pattern Recognition Lab we are involved in research projects and supervise student and master theses. This connection allows to perform feasibility studies in the latest pattern recognition technologies for both medical and industrial applications.

Chimaera can therefore establish the link between research and a final commercial product in the field of pattern recognition.

Features

We have held courses and exercises at the university that cover pattern recognition topics ranging from the very basics to recently developed advanced concepts. The techniques can be widely used within industrial and medical imaging applications.

  • 2-D, 3-D and 4-D signal analysis
  • Data filtering techniques
  • Feature extraction
  • Statistical estimation
  • Statistical models
  • Linear and non-linear classification systems
  • Boosting techniques
  • Multi-sensor data fusion
  • GPU-accelerated processing
  • Pattern analysis for real-time applications
  • Collaboration with the Pattern Recognition Lab
  • Link between academia and industry
  • Medical and industrial applications

Examples

Kidney Segmentation using Statistical Shape Models

Active shape models (ASMs) are an example for statistical modeling. ASMs are widely used for image segmentation due to their inherent statistical regularization with prior knowledge. This makes them robust to leaking problems if adjacent image structures are not clearly delineated from each other. The core of ASMs is a statistical model that is gained from the variations in the training data. The statistical model can then be used for the classification of structures or organs of interest.

There is, however, a difference in the way statistical models are generated. In the left image, the colors encode the distances between a gold standard kidney segmentation that was not in the training data (shown as surface) and the result of two shape model segmentations (color on the surface). The color encodes zero distance in red to maximal distance in blue.

The left result has been computed using a shape model that was generated with our registration techniques, whereas the right result was computed with a state-of-the-art technique for ASMs. It shows larger distances in the final segmentation.

A typical segmentation result using a statistical shape model is depicted on the right for the case of kidney segmentation.

In all ASM techniques, there is a trade-off between the possibility to accurately detect the boundaries of an organ shape and the model regularization. The more weight is assigned to the training data, the closer the final segmentation will stay within its trained limits.

For a commercial product, Chimaera can provide the engineering that is necessary to achieve a solution that is robust for various images without cumbersome parameter adjustments.

The images have been taken from: Spiegel, Martin; Hahn, Dieter; Daum, Volker; Wasza, Jakob; Hornegger, Joachim. Segmentation of kidneys using a new active shape model generation technique based on non-rigid image registration. In: Computerized Medical Imaging and Graphics 33 (2009) No. 1 pp. 29-39.