Publication detail

Artificial Image Synthesis and Data Augmentation for Deep Learning Segmentation of Phase Contrast Images for Biomarker Discovery in Cancer Research

MÍVALT, F. KROMP, F. LAZIC, D. OSTALECKI, C. AMBROS, I. TASCHNER-MANDL, S. AMBROS, P.

Original Title

Artificial Image Synthesis and Data Augmentation for Deep Learning Segmentation of Phase Contrast Images for Biomarker Discovery in Cancer Research

Type

presentation, poster

Language

English

Original Abstract

Deep learning (DL) algorithms are achieving or even surpassing human-level performance in tasks like image classification or segmentation. Due to their fast development and exceptional performance, DL algorithms were introduced in various life science domains such as biomedical imaging, bioinformatics or computational biology. However, the outcome of these algorithms on unseen data highly depends on the quality of the training dataset. Thus, there is a need for manual data annotation which is a lengthy and time-consuming process, especially in the field of cell imaging. We hereby propose a technique to accelerate data annotation by using synthetic phasecontrast images to train deep learning algorithms for cell segmentation. Uniformity of the statistical data distribution, specific image artefacts modelling or representability were defined, and the method was designed and implemented. The feasibility of the proposed method was demonstrated by training the Mask R-CNN model for instanceaware segmentation using only synthetic images. The evaluation was performed on synthetic and real images as well. The F1 score for real images was 95.68% emphasising the advantage of using synthetic images for the training of DL algorithms. The segmentation of phase-contrast images enables the subsequent combined analysis of corresponding fluorescent microscopy images on the single cell and sub-cellular level and support biomarker discovery in cancer research.

Keywords

deep learning, cell segmentation, phase-contrast, data augmentation, artificial data generation

Authors

MÍVALT, F.; KROMP, F.; LAZIC, D.; OSTALECKI, C.; AMBROS, I.; TASCHNER-MANDL, S.; AMBROS, P.

Released

13. 6. 2019

Publisher

Young Scientist Association - PhD Symposium

Location

Vienna

Pages count

1

URL

BibTex

@misc{BUT164888,
  author="Filip {Mívalt} and Florian {Kromp} and Daria {Lazic} and Christian {Ostalecki} and Inge {Ambros} and Peter {Ambros} and Sabine {Taschner-Mandl}",
  title="Artificial Image Synthesis and Data Augmentation for Deep Learning Segmentation of Phase Contrast Images for Biomarker Discovery in Cancer Research",
  year="2019",
  pages="1",
  publisher="Young Scientist Association - PhD Symposium",
  address="Vienna",
  doi="10.13140/RG.2.2.12694.50243",
  url="https://www.researchgate.net/publication/335363685_Artificial_Image_Synthesis_and_Data_Augmentation_for_Deep_Learning_Segmentation_of_Phase_Contrast_Images_for_Biomarker_Discovery_in_Cancer_Research",
  note="presentation, poster"
}