CycleGAN With an Improved Loss Function for Cell Detection Using Partly Labeled Images.

Area of Science:

• Biomedical imaging
• Computer vision
• Machine learning

Background:

• Object detection is crucial in biomedicine but challenged by data quality.
• Incomplete or inaccurate labeling, especially in cell detection, hinders model performance.
• Existing public datasets often lack the complexity of real-world scenarios.

Purpose of the Study:

• To develop a data augmentation algorithm for generating fully labeled cell image data from incomplete or inaccurately labeled sources.
• To address the weak annotation problem in cell detection.
• To improve the performance of object detection models in biomedical imaging.

Main Methods:

• A novel data augmentation algorithm is proposed.
• Labeled objects are extracted from raw cell images, preserving positional information.
• A modified cycle-consistent adversarial network framework generates complete labeled data, including objects and backgrounds.

Main Results:

• The proposed method successfully generates fully labeled cell image data.
• Experiments on the Blood Cell Classification Dataset (BCCD) demonstrate effectiveness.
• The approach addresses the weak annotation problem in object detection.

Conclusions:

• The developed data augmentation technique effectively overcomes challenges posed by incomplete and inaccurate data labeling.
• The method significantly improves object detection performance in the biomedical field, specifically for cell analysis.
• This work offers a valuable solution for enhancing machine learning models trained on imperfect biomedical image datasets.