Combining Deep Learning with Handcrafted Features for Cell Nuclei Segmentation^.

Area of Science:

• Cell Biology
• Biomedical Imaging
• Computational Biology

Background:

• Accurate cell nuclei segmentation is crucial for analyzing cellular characteristics like shape, size, chromatin texture, and DNA content.
• Applications include cell tracking, counting, and classification in microscopy.
• Deep learning approaches have shown promise but can be further improved with domain knowledge.

Purpose of the Study:

• To enhance a deep learning-based nuclei segmentation method by incorporating handcrafted features.
• To leverage domain knowledge about the approximately round shape of cell nuclei.
• To improve segmentation accuracy and convergence speed.

Main Methods:

• Extended a previously proposed deep learning approach for nuclei segmentation.
• Introduced handcrafted features representing the roundness of nuclei, specifically gradient convergence.
• Computed a gradient convergence map as an additional input channel for the Convolutional Neural Network (CNN).
• Applied the enhanced method to DAPI-stained cell microscopy images.

Main Results:

• The enhanced method decreased miss-detections compared to the previous approach.
• The F1-Score for nuclei segmentation was increased.
• Faster convergence of the deep learning model was observed when handcrafted features were combined with deep learning.
• The method demonstrated improved performance on DAPI-stained cell microscopy images.

Conclusions:

• Integrating handcrafted features, such as gradient convergence, significantly improves deep learning-based cell nuclei segmentation.
• The combined approach offers higher accuracy (increased F1-Score) and efficiency (faster convergence).
• This method provides a more robust tool for quantitative analysis of cell nuclei in fluorescence microscopy.