GRUU-Net: Integrated convolutional and gated recurrent neural network for cell segmentation.

Area of Science:

• Computer Vision
• Bioimaging Analysis
• Machine Learning

Background:

• Cell segmentation in microscopy is crucial but challenging.
• Deep learning, particularly convolutional neural networks (CNNs), has advanced computer vision tasks.
• Recurrent neural networks (RNNs) are less common for segmentation but offer unique capabilities.

Purpose of the Study:

• To develop a novel deep learning method for improved cell segmentation.
• To leverage the strengths of both CNNs and gated RNNs across multiple image scales.
• To enhance segmentation robustness and accuracy using a new focal loss function.

Main Methods:

• Integration of CNNs and gated RNNs for multi-scale image analysis.
• Introduction of a novel focal loss function for robust training.
• Implementation of a distributed training scheme for optimized performance.
• Application to glioblastoma cell nuclei and benchmarking on 22 Cell Tracking Challenge datasets.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art techniques.
• Quantitative comparisons highlight the effectiveness of the integrated network architecture.
• Analysis provides insights into the impact of extensions on training and inference.

Conclusions:

• The combined CNN-RNN approach offers significant improvements in cell segmentation.
• The novel focal loss and distributed training enhance method robustness and efficiency.
• The method is validated across diverse microscopy datasets, showing broad applicability.