Automated Early Cancer Screening Based on Kernel Method
Baochuan Pang, Yiming Lu, Duanquan Xu · 2008
An automated method that detects early cancerous specimens based on image analysis is described. After acquisition and noise reduction, the microscope images are segmented into individual cell nucleus, from which the feature vectors of nucleus are calculated. The dimensionality of the feature vectors is then reduced using a method combing F-Score and random forest algorithms. The types of the cell nucleus are identified by a classifier based on a non-linear kernel method, and the diagnosis is made on the basis of the statistics. The method was experimented on a data set of 25,000 cell nucleus instances extracted from 5,000 images of 50 specimens. When tested with 5-fold cross-validation algorithm, this early cancer detecting method resulted in the correct classification of over 97% of the cell nucleus. All cancerous positive specimens were successfully detected in the experiment.