Auto-CSC: A Transfer Learning Based Automatic Cell Segmentation and Count Framework.

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence in Medicine

Background:

• Accurate cell segmentation and counting are crucial for disease diagnosis, particularly in hematology.
• Deep learning models, like convolutional neural networks, show promise for image segmentation but require substantial manual annotation.
• Manual annotation is labor-intensive, costly, and susceptible to human error.

Purpose of the Study:

• To develop a novel framework for automated cell segmentation and counting that minimizes the requirement for manually annotated cell images.
• To improve the efficiency and reduce the cost associated with preparing training data for cell analysis.

Main Methods:

• Generated synthetic cell image labels using traditional algorithms on single-kind cell images.
• Trained a segmentation model (U-Net) using these synthetically labeled images in stages, updating parameters with different cell types.
• Transferred the pretrained model to segment mixed cell images using a small dataset of manually labeled mixed cell images.

Main Results:

• The proposed method achieved cell segmentation and counting performance equivalent to models trained on large, manually annotated datasets.
• The framework effectively segmented and counted cells in mixed populations after pretraining on synthetically labeled single-kind cell images.
• Validation demonstrated the method's effectiveness in reducing annotation burden without compromising accuracy.

Conclusions:

• The novel framework significantly reduces the need for manual annotations in cell segmentation and counting tasks.
• This approach offers a cost-effective and efficient alternative for medical image analysis, maintaining high performance.
• The study validates the transfer learning strategy for cell segmentation using synthetically generated labels.