Reducing Manual Annotation Costs for Cell Segmentation by Upgrading Low-Quality Annotations.

Area of Science:

• Computational Biology
• Medical Imaging Analysis
• Artificial Intelligence in Medicine

Background:

• Deep learning for cell segmentation demands extensive, high-quality annotated datasets, which are costly and time-consuming to produce.
• High annotation costs can be a significant barrier to developing accurate cell segmentation models.

Purpose of the Study:

• To develop a deep learning method for upgrading low-quality cell image annotations, thereby reducing annotation time and cost.
• To evaluate the effectiveness of the proposed method in correcting various types of annotation errors (omission, inclusion, bias).
• To demonstrate the utility of upgraded annotations in training improved cell segmentation networks.

Main Methods:

• A convolutional neural network was trained on a small, high-quality dataset to learn how to upgrade lower-quality annotations.
• The method was tested on annotations with simulated omission, inclusion, and bias errors.
• The performance was quantified using Dice similarity coefficient against ground-truth annotations.

Main Results:

• The proposed method successfully upgraded annotations with high error levels, achieving Dice similarity scores up to 0.9.
• Training cell segmentation networks on a combination of well-annotated and upgraded data resulted in better performance than using only the well-annotated set.
• A use case demonstrated successful quality improvement for predictions from a network trained on only 10 annotated samples.

Conclusions:

• Deep learning can effectively upgrade low-quality cell image annotations, significantly mitigating annotation costs.
• Enlarging small, high-quality datasets with upgraded annotations improves the performance of cell segmentation models.
• This approach offers a practical solution for enhancing cell image analysis in resource-constrained settings.