Domain Adapted Multitask Learning for Segmenting Amoeboid Cells in Microscopy.

Area of Science:

• Computational Biology
• Biomedical Image Analysis
• Machine Learning

Background:

• Accurate segmentation of amoeboid cells in microscopy is crucial for biological research.
• Current Convolutional Neural Network (CNN)-based segmentation methods often struggle with ad-hoc hyperparameter estimation.
• Addressing hyperparameter tuning is essential for improving the robustness of CNN segmentation models.

Purpose of the Study:

• To develop a robust method for segmenting amoeboid cells in microscopy images.
• To analytically estimate model hyperparameters during the training of CNNs for segmentation.
• To improve cell segmentation accuracy in both brightfield and fluorescence microscopy images.

Main Methods:

• A novel framework combining multi-task learning and unsupervised domain adaptation was proposed.
• A min-max formulation of the segmentation cost function was used for analytical hyperparameter estimation.
• The end-to-end framework simultaneously learns CNN weights and estimates hyperparameters.

Main Results:

• The proposed method demonstrated effectiveness in segmenting clustered cells from low-contrast brightfield images.
• Deep domain adaptation enabled segmentation of fluorescent cells without pixel-level re-annotation.
• Quantitative results showed at least 15% improvement in brightfield and 10% in fluorescence image segmentation compared to supervised methods.

Conclusions:

• The proposed method offers a consolidated approach for robust cell segmentation in microscopy.
• Analytical hyperparameter estimation enhances the performance of CNN-based segmentation models.
• The technique significantly improves segmentation accuracy across different imaging modalities.