gms | German Medical Science

67. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e. V. (GMDS), 13. Jahreskongress der Technologie- und Methodenplattform für die vernetzte medizinische Forschung e. V. (TMF)

21.08. - 25.08.2022, online

MRI Brain Image Segmentation with Machine Learning for Mice and Rats: A Preclinical Application

Meeting Abstract

  • Lucas Plagwitz - Westfälische Wilhelms-Universität Münster, Institut für Medizinische Informatik, Münster, Germany
  • Bruno Pradier - Klinik für Radiologie, Translational Research Imaging Center, Universitätsklinikum Münster, Münster, Germany
  • Catharina van Alen - Westfälische Wilhelms-Universität Münster, Institut für Medizinische Informatik, Münster, Germany
  • Julian Varghese - Westfälische Wilhelms-Universität Münster, Institut für Medizinische Informatik, Münster, Germany

Deutsche Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie. 67. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e. V. (GMDS), 13. Jahreskongress der Technologie- und Methodenplattform für die vernetzte medizinische Forschung e.V. (TMF). sine loco [digital], 21.-25.08.2022. Düsseldorf: German Medical Science GMS Publishing House; 2022. DocAbstr. 135

doi: 10.3205/22gmds052, urn:nbn:de:0183-22gmds0521

Published: August 19, 2022

© 2022 Plagwitz et al.
This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0 License. See license information at http://creativecommons.org/licenses/by/4.0/.


Outline

Text

Introduction: In radiology, the preparation of magnetic resonance imaging (MRI) data is an essential step for subsequent analyses. Although it has been shown that modern machine learning algorithms can assist radiologists in MRI-specific preprocessing tasks [1], they have often not yet arrived in preclinical practice. One instance of this is brain masking, which focuses measurement data within the brain while eliminating skin, bone, and artifacts outside the brain [2]. Existing machine learning algorithms for brain masking have mostly been trained using human brain image data. Thus, for specific data of laboratory animals (e.g., mice and rats) an adjusted model has to be trained. In this context, an estimate of effort is needed: (1) the amount of data that would be required for mouse-based brain masking and (2) the robustness of such algorithms with respect to different input data, e.g., whether a mouse data-trained model is able to segment rat brains.

Methods: For masking of three-dimensional (3D) functional brain images, we implemented the deep learning architecture U-Net, which is an accomplished method in the field of image segmentation [3]. While 3D models have already been established, we considered 2D slices to generate as many training samples as possible with few MRI measurements. The model was trained with varying quantities of animals and cross-validated to assess performance. Each animal was either in the training or test set. These performances were compared using the Dice score (DS), which is a common metric in medical image segmentation analysis [4]. Masks drawn by imaging experts on 67 animals (60 mice and 7 rats) formed the basis for training and testing.

Results: While the model trained on 3 mice (54 slices – 0.935 DS) segmented correctly in most instances, an increase to 30 mice (540 slices – 0.961 DS) was sufficient for the model to learn the concept of masking mouse brains. The DS hardly improved with additional mice.

The model (trained with mice) was directly transferable to image data of rats and showed similar behavior. Overall, the performances were lower (e.g., 30 mice – 0.804 DS) but did not increase when the training set included more than 30 mice. The performance did however improve significantly when adding 5 rats to the training set (DS 0.905).

Discussion: The analysis showed that a small amount of mice is sufficient for the model to learn the concept of brain masking, which can be explained by the simplification down to two dimensions and the resulting enlargement of the training set. Above a threshold of 30 mice, the DS did not increase, which is consistent with the non-existence of unique mask boundaries. Furthermore, the segmentation is shown to be robust when transferred to image data of rats. Nevertheless, the model clearly benefits from adding rat data to the training set.

Conclusion: This analysis enabled the introduction of machine learning into everyday MRI preprocessing. In this way, the theory of U-Net architecture enters our preclinical routine of radiology and saves a significant amount of time.

Publicly available implementation: https://github.com/lucasplagwitz/rm_masking.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.


References

1.
Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts HJWL. Artificial intelligence in radiology. Nature Reviews. Cancer. 2018; 18(8):500-510.
2.
Despotović I, Goossens B, Philips W. MRI segmentation of the human brain: challenges, methods, and applications. Computational and mathematical methods in medicine. 2015;2015:450341.
3.
Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. Medical Image Computing and Computer-Assisted Intervention – MICCAI. 2015;9351:234-241.
4.
Eelbode T, Bertels J, Berman M, et al. Optimization for Medical Image Segmentation: Theory and Practice When Evaluating With Dice Score or Jaccard Index. IEEE Trans Medical Imaging. 2020;39(11):3679-3690.