Identifying Partial Mouse Brain Microscopy Images from the Allen Reference Atlas Using a Contrastively Learned Semantic Space
Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review
Standard
Identifying Partial Mouse Brain Microscopy Images from the Allen Reference Atlas Using a Contrastively Learned Semantic Space. / Antanavicius, Justinas; Leiras, Roberto; Selvan, Raghavendra.
Biomedical Image Registration - 10th International Workshop, WBIR 2022, Proceedings. ed. / Alessa Hering; Julia Schnabel; Miaomiao Zhang; Enzo Ferrante; Mattias Heinrich; Daniel Rueckert. 1. ed. Springer Science and Business Media Deutschland GmbH, 2022. p. 166-176 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 13386 LNCS).Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review
Harvard
APA
Vancouver
Author
Bibtex
}
RIS
TY - GEN
T1 - Identifying Partial Mouse Brain Microscopy Images from the Allen Reference Atlas Using a Contrastively Learned Semantic Space
AU - Antanavicius, Justinas
AU - Leiras, Roberto
AU - Selvan, Raghavendra
N1 - Publisher Copyright: © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2022
Y1 - 2022
N2 - Registering mouse brain microscopy images to a reference atlas is crucial to determine the locations of anatomical structures in the brain, which is an essential step for understanding the function of brain circuits. Most existing registration pipelines assume the identity of the reference plate – to which the image slice is to be registered – is known beforehand. This might not always be the case due to three main challenges in microscopy image data: missing image regions (partial data), different cutting angles compared to the atlas plates and a large number of high-resolution images to be identified. Manual identification of reference plates as an initial step requires highly experienced personnel and can be biased, tedious and resource intensive. On the other hand, registering images to all atlas plates can be slow, limiting the application of automated registration methods when dealing with high-resolution image data. This work proposes to perform the image identification by learning a low-dimensional space that captures the similarity between microscopy images and the reference atlas plates. We employ Convolutional Neural Networks (CNNs), in the Siamese network configuration, to first obtain low-dimensional embeddings of microscopy image data and atlas plates. These embeddings are contrasted with positive and negative examples in order to learn a semantically meaningful space that can be used for identifying corresponding 2D atlas plates. At inference, atlas plates that are closest to the microscopy image data in the learned embedding space are presented as candidates for registration. Our method achieved TOP-3 and TOP-5 accuracy of 83.3% and 100%, respectively, compared to the SimpleElastix-based baseline which obtained 25% in both the Top-3 and Top-5 accuracy (Source code is available at https://github.com/Justinas256/2d-mouse-brain-identification ).
AB - Registering mouse brain microscopy images to a reference atlas is crucial to determine the locations of anatomical structures in the brain, which is an essential step for understanding the function of brain circuits. Most existing registration pipelines assume the identity of the reference plate – to which the image slice is to be registered – is known beforehand. This might not always be the case due to three main challenges in microscopy image data: missing image regions (partial data), different cutting angles compared to the atlas plates and a large number of high-resolution images to be identified. Manual identification of reference plates as an initial step requires highly experienced personnel and can be biased, tedious and resource intensive. On the other hand, registering images to all atlas plates can be slow, limiting the application of automated registration methods when dealing with high-resolution image data. This work proposes to perform the image identification by learning a low-dimensional space that captures the similarity between microscopy images and the reference atlas plates. We employ Convolutional Neural Networks (CNNs), in the Siamese network configuration, to first obtain low-dimensional embeddings of microscopy image data and atlas plates. These embeddings are contrasted with positive and negative examples in order to learn a semantically meaningful space that can be used for identifying corresponding 2D atlas plates. At inference, atlas plates that are closest to the microscopy image data in the learned embedding space are presented as candidates for registration. Our method achieved TOP-3 and TOP-5 accuracy of 83.3% and 100%, respectively, compared to the SimpleElastix-based baseline which obtained 25% in both the Top-3 and Top-5 accuracy (Source code is available at https://github.com/Justinas256/2d-mouse-brain-identification ).
KW - Deep learning
KW - Image registration
KW - Mouse brain
KW - Partial data
U2 - 10.1007/978-3-031-11203-4_18
DO - 10.1007/978-3-031-11203-4_18
M3 - Article in proceedings
AN - SCOPUS:85135077575
SN - 9783031112027
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 166
EP - 176
BT - Biomedical Image Registration - 10th International Workshop, WBIR 2022, Proceedings
A2 - Hering, Alessa
A2 - Schnabel, Julia
A2 - Zhang, Miaomiao
A2 - Ferrante, Enzo
A2 - Heinrich, Mattias
A2 - Rueckert, Daniel
PB - Springer Science and Business Media Deutschland GmbH
T2 - 10th International Workshop on Biomedical Image Registration, WBIR 2020
Y2 - 10 July 2022 through 12 July 2022
ER -
ID: 315633381