| Bibtype | Inproceedings |
|---|---|
| Bibkey | Siedhoff/etal/2011a |
| Author | Siedhoff, Dominic and Weichert, Frank and Libuschewski, Pascal and Timm, Constantin |
| Title | Detection and Classification of Nano-Objects in Biosensor Data |
| Booktitle | Microscopic Image Analysis with Applications in Biology |
| Abstract | Preventing viral infections from spreading quickly in a heavily connected world demands for reliable diagnostic methods providing results promptly. The PAMONO biosensor (Plasmon Assisted Microscopy Of Nano-Objects) is a novel technique capable of attaining these properties.
In this paper, a processing pipeline is proposed for analyzing PAMONO sensor data. Firstly, virus candidate pixels are detected in the sensor output by matching their series of measured intensities over time to characteristic patterns. This is achieved in a fully parallel GPGPU approach. Each spatially coherent set of matching pixels defines a candidate object, represented as a polygon. The overall set of polygons encompasses true virus polygons as well as several types of false detections. The polygons are classified based on their shape, separating the viruses from false detections. With regard to this step, the suitability of different geometrical features (form-factors) and classification methods is explored. A set of classifiers consisting of Naive Bayes, RIPPER rule induction, C4.5 Decision Trees, k-Nearest-Neighbor and Support Vector Machines (SVM) is applied to a two- and to a multi-class formulation of the classification problem. Training and searching for robust optimal parameterizations of the parametric classifiers is achieved in an offline step, using evolutionary optimization. Furthermore, an evolutionary feature-selection is conducted for all classifiers. The classification performances are evaluated for different types of polystyrene nano-particles. With regard to the diverse nature of the encountered artifacts, a one-class SVM approach, learning from positive examples only, is an attractive option. As will be shown, it can compete with the two- and multi-class approaches if training and test data originate from the same type of nano-particles. |
| Year | 2011 |
| Projekt | SFB876-B2 |