CACs Recognition of FISH Images Based on Adaptive Mean Teacher Semi-supervised Learning with Domain-Knowledge Pseudo...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Oncology

Background:

• Circulating abnormal cells (CACs) are vital biomarkers for lung cancer diagnosis and early screening.
• Deep learning algorithms offer potential for CACs detection but require extensive data labeling.
• Reducing reliance on labeled data is crucial for practical deep learning applications in cancer diagnostics.

Purpose of the Study:

• To develop a semi-supervised learning algorithm for enhanced detection of circulating abnormal cells (CACs) in lung cancer.
• To improve cell segmentation and signal point detection for CACs identification using minimal labeled data.
• To reduce the dependency on large, manually labeled datasets in deep learning-based cancer diagnostic systems.

Main Methods:

• Employed a semi-supervised approach combining self-training and Mean Teacher methods for cell segmentation.
• Developed an Adaptive Mean Teacher approach to optimize semi-supervised cell segmentation.
• Implemented an end-to-end semi-supervised signal point detection algorithm using Adaptive Mean Teacher and Domain-Knowledge Pseudo Labels.

Main Results:

• The semi-supervised method demonstrated effectiveness in cell segmentation and signal point detection tasks.
• Achieved significant performance in the final CACs detection task with limited labeled data (2%, 5%, 10%).
• The proposed approach successfully reduced the need for extensive data labeling while maintaining high performance.

Conclusions:

• Semi-supervised learning effectively enhances CACs detection for lung cancer screening.
• The developed Adaptive Mean Teacher and Domain-Knowledge Pseudo Label strategies improve deep learning model performance with less labeled data.
• This approach offers a viable solution for practical, data-efficient deep learning in early lung cancer diagnosis.