A Novel Graph-Based Approach for Seriation of Mouse Brain Cross-Section from Images

• Published in: Pattern Recognition and Image Analysis pp. 461 - 471
• Main Authors: Sarbazvatan, S., Ventura, R., Esteves, F. F., Lima, S. Q., Sanches, J. M.
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: (dis)similarity matrix; Graph; HOG; Mouse brain; Seriation
• ISBN: 303031331X; 9783030313319
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-31332-6_40

This paper addresses the problem of automatic seriation of mouse brain cross-sections stained with green-florescence protein (GFP). This is fundamental for the neuroscience community to help in the processing and analyzing the huge amount of experimental data. It is also a challenging problem since, during the manual procedure of cutting the brains and acquiring hundreds of images, the human operator can unwittingly change its natural sequence, loose data, induce large morphological distortions, or introduce artifacts. Most image seriation methods are two-step: firstly, a distance matrix is obtained from image processing, and secondly, the optimal seriation method is determined for this matrix. However, these methods are very sensitive to noise, distortion, and missing data, since the optimal solution for the matrix does not match the true seriation. Instead, we propose a graph-based method where the images are iteratively revisited and the image similarity information is refined, until a linear graph representing the seriation is obtained. This similarity information is based on Histogram Oriented Gradient (HOG) features, computed from random locations at the images in each iteration/revisitation. Experimental results based on both synthetic and real data are used to validate and illustrate the application of the method. It is showed that the proposed method outperforms the other state-of-the-art methods used for comparison purposes in this specific type of data.