Learning to distinguish cerebral vasculature data from mechanical chatter in India-ink images acquired using knife-edge scanning microscopy.

We introduce a simple, yet effective, procedure for accurate classification of connected components embedded in biological images. In our method, a training set is generated from user-delineated features of manually-labeled examples; we subsequently train a classifier using the resultant training set. The overall process is described using imaging data acquired from an India-ink perfused C57BL/6J mouse brain using Knife Edge Scanning Microscopy. We illustrate the procedure through segmentation of cerebral vasculature structures from mechanical noise using trained classifiers. The features extracted by our procedure show high discriminatory power between classes; the classifiers (linear SVM, Gaussian SVM, and GentleBoost decision tree ensemble) trained using these features achieved high performance: F1-scores reported for linear SVM, Gaussian SVM, and GentleBoost decision tree ensemble were 0.963, 0.956, and 0.963 respectively.