Effective pseudo-labeling based on heatmap for unsupervised domain adaptation in cell detection.

Area of Science:

• Biomedical research
• Computational biology
• Image analysis

Background:

• Cell detection is crucial in biomedical research, with deep learning enhancing performance.
• Domain shift, where models trained on one data set fail on another, hinders cell detection accuracy due to varying cell culture conditions (e.g., shape, density).

Purpose of the Study:

• To propose an unsupervised domain adaptation method for robust cell detection.
• To address the domain shift problem in cell detection caused by variations in experimental conditions.

Main Methods:

• Developed an unsupervised domain adaptation technique utilizing pseudo-cell-position heatmaps.
• Implemented selective pseudo-labeling by regenerating heatmaps to ensure a Gaussian distribution around cell centroids.
• Employed uncertainty and curriculum learning for selecting confident pseudo-labels.

Main Results:

• The proposed method significantly improved cell detection performance across diverse experimental conditions.
• Demonstrated superior performance compared to existing domain adaptation methods in numerous experiments.
• Successfully adapted cell detection models to new target domains without requiring labeled data from those domains.

Conclusions:

• The novel pseudo-cell-position heatmap approach effectively mitigates domain shift in cell detection.
• This method offers a promising solution for accurate and reliable cell detection in varied biomedical research settings.
• Unsupervised domain adaptation is a viable strategy for enhancing the generalizability of deep learning models in biological image analysis.