Unsupervised Annotation Transfer in Phase-Contrast Microscopy Using a CycleGAN

Area of Science:

• Microscopy imaging
• Deep learning
• Computational biology

Background:

• Domain shift in microscopy imaging hinders deep learning model deployment.
• Manual annotation of new datasets is costly and time-consuming.

Purpose of the Study:

• To develop a Cycle-Consistent Generative Adversarial Network (CycleGAN)-based framework for annotation transfer.
• To adapt a labeled B16BL6 cell microscopy domain to a HeLa target domain without extensive manual annotation.

Main Methods:

• Utilized a detection-oriented CycleGAN for image translation from B16BL6 to HeLa domains.
• Retained bounding-box annotations during image translation.
• Trained YOLOv8x object detectors on source-only and translated datasets.
• Evaluated performance using standard object detection metrics and unsupervised proxy metrics.

Main Results:

• The CycleGAN-translated configuration achieved superior performance (mAP@0.50: 0.186) compared to baselines (mAP@0.50: 0.027 and 0.009).
• Achieved highest precision (0.244), recall (0.353), and F1-score (0.288) on the target domain.
• Demonstrated improved detector generalization and reduced need for manual annotation.

Conclusions:

• Source-to-target image translation effectively addresses domain shift challenges in microscopy.
• The proposed CycleGAN framework significantly enhances deep learning model adaptability across different experimental setups.
• Reduces the burden of manual annotation for new microscopy datasets.