Data Discernment for Affordable Training in Medical Image Segmentation.

Area of Science:

• Medical image analysis
• Deep learning
• Computer vision

Background:

• High-quality training data is crucial but expensive for medical image segmentation using deep neural networks.
• External data sources, like crowd-sourcing, offer a cost-effective alternative but often have different data distributions.

Purpose of the Study:

• To develop a method for training deep neural networks using external data despite distribution discrepancies.
• To enhance the utility of crowd-sourced or other external data for medical image segmentation tasks.

Main Methods:

• Proposing a data discernment technique to assign importance weights to external data points.
• Developing an iterative algorithm that alternately estimates weights and updates the network, formulated as a constrained nonlinear programming problem.
• Estimating weights based on distribution discrepancy and enforcing constraints for effective network learning.

Main Results:

• Demonstrated the ability of deep neural networks to mine valuable knowledge from external data with differing distributions.
• Showcased improved performance in abdominal CT and cervical smear image segmentation tasks.
• Validated the effectiveness of the data discernment algorithm through extensive experiments on public datasets.

Conclusions:

• Data discernment enables effective utilization of cheaper, external datasets for medical image segmentation.
• The proposed iterative algorithm successfully enhances informative external data contributions while suppressing irrelevant or harmful data.
• This approach offers a viable solution to the challenge of data scarcity in medical deep learning.