Unsupervised learning for biomedical applications

With the growth of application of computers in the generation and analysis of biomedical data, a variety of computerized methods and algorithms have been proposed to optimize the process of acquisition and analysis of the data. Although advanced computerized techniques have provided the means for more precise diagnosis, the interpretation of the recorded data in some cases is an issue due to the large amount of the data or complexity of it. While most of the existing work in the literature consider supervised techniques for analysis of the collected data, the use of unsupervised techniques in the area of analysis and classification of biomedical signals is relatively unexplored compared to supervised approaches. In general, the investigation of application of unsupervised techniques for analysis of biomedical signals can be worthwhile from different view points. In some cases, biomedical databases tend to contain a large amount of data. Genomic databases or pathological speech databases are examples of this kind. The development of any supervised method for analysis of such databases requires precise manual labeling of the data, which can be extremely costly. However, the use of an unsupervised classifier can be beneficial to accelerate the process and to acquire information about the structure of the dataset. In addition, the characteristics of the collected biomedical data can be affected by the recording process. In this work application of unsupervised learning in two biomedical signal processing problems is investigated. In the first problem, fuzzy C-means clustering has been used in design of a computer aided diagnosis method for detection of abnormalities in small bowel capsule endoscope images. The performance of the system shows an accuracy of 76which is an acceptable rate for an unsupervised method. In the second case, self organizing tree maps (SOTM) has been applied to audio signal classification for hearing aids. An accuracy of 96% was achieved for discrimination of human voice from the environmental noise, which is one the major classification scenarios for hearing aid applications.