Feature-driven Landmark Detection in Gene Expression Mouse Brain Images using Texture Analysis

Background: To better understand the development and function of the mammalian brain, researchers collect a large number of gene expression patterns throughout the mouse brain using in-situ hybridization technology, in which 20um tissue slices are stained with probes for desired genes and imaged at high resolution (3.3um per pixel). The overall aim of this project is to develop a framework to automatically measure the gene expressions in each anatomical region and at different developmental stages of the mouse brain. Our specific aim is to develop a method to detect anatomical landmarks that can be used to guide the fitting of a standard deformable atlas to any given expression image. Methods: Given images of sagittal slices of the mouse brain, we selected consistent landmarks and extracted texture information and statistical parameters like the mean intensity and standard deviation from a region around them. Discriminant Analysis with Wilks lambda was used to select the best features. The training set of images was selected using a forward selection procedure. This selected training set was used to model the Support Vector Machine (SVM) based classifier using the reduced features. A best match was found around the landmark point in the testing images using this SVM. Results: In the first set of tests conducted, 4 landmarks were accurately detected in 76% of the images. Conclusion: We have developed an automated landmark detection method with very encouraging results.