Related Experiment Video
Updated: Dec 28, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
941
Prediction of Sequential Organelles Localization under Imbalance using A Balanced Deep U-Net.
Novanto Yudistira1,2, Muthusubash Kavitha3, Takeshi Itabashi4,5,6
1Hiroshima University, Department of Information Engineering, Higashi Hiroshima, 739-8521, Japan. yudistira@ub.ac.id.
Scientific Reports
|February 16, 2020
Summary
A novel deep learning approach using a balanced regularized weighted compound dice loss (RWCDL) network improves the automated identification of cellular structures in red algae during mitosis. This method enhances organelle segmentation, particularly for small structures like peroxisomes and nuclei.
Area of Science:
- Cell Biology
- Bioimaging
- Machine Learning
Background:
- Accurate segmentation of cellular organelles in 3D microscopy images is crucial for understanding cell function.
- Challenges exist in automated techniques due to class imbalance, especially during mitosis in unicellular organisms like Cyanidioschyzon merolae.
- Existing methods struggle with reliably identifying smaller or less frequent organelles.
Purpose of the Study:
- To develop a robust automated technique for segmenting cellular organelles in Cyanidioschyzon merolae during mitosis.
- To address the issue of class imbalance in organelle segmentation using a novel deep learning approach.
- To improve the localization accuracy of small and obscure organelles, such as peroxisomes and nuclei.
Main Methods:
- Proposed a balanced deep regularized weighted compound dice loss (RWCDL) network, extending a Unet-like convolutional neural network (CNN) architecture.
- Introduced two new loss functions: compound dice (CD) and RWCD, incorporating multi-class dice variants and a weighting mechanism.
- Focused on maximizing the weights of peroxisomes and nuclei among five cellular structure classes.
Main Results:
- The RWCDL network achieved significantly higher area under the curve (AUC) values, improving segmentation of small organelles by up to 30% compared to standard methods (MSE, DL).
- Demonstrated reliable identification of cellular structures even with obscure cell contours.
- Validated the approach on three large-scale mitotic cycle datasets with varying organelle occurrences.
Conclusions:
- The balanced deep RWCDL approach offers a reliable solution for automated organelle segmentation in challenging biological imaging scenarios.
- This technique is valuable for biologists seeking to accurately correlate cell behavior with organelle structures during mitosis.
- The proposed method enhances the study of primitive unicellular red algae by improving the precision of cellular structure analysis.
