Deep learning segmentation of endothelial cell images using an active learning paradigm with guided label corrections

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate corneal endothelial cell (EC) segmentation is crucial for assessing post-keratoplasty outcomes.
• Automated segmentation methods often require manual correction, which is time-consuming.
• Deep learning models show promise but need high-quality training data.

Purpose of the Study:

• To develop Guided Correction Software for efficient manual editing of automated EC segmentations.
• To apply this software within an active learning framework for analyzing post-keratoplasty EC images.
• To improve the performance and generalization of deep learning models for EC segmentation.

Main Methods:

• A U-Net model generated initial EC segmentations on post-keratoplasty images.
• Guided Correction Software was used to manually refine these segmentations, creating corrected labels.
• U-Net and DeepLabV3+ models were trained using both uncorrected and corrected segmentation labels.
• Performance was evaluated based on segmentation accuracy, cell identification, and endothelial cell density (ECD) estimation.

Main Results:

• Training deep learning models with corrected segmentations significantly improved performance, reducing over- and under-segmentations.
• Models trained on corrected data achieved accurate ECD predictions comparable to ground truth.
• The Guided Correction Software and semi-automated workflow reduced image segmentation time by over 80% (from 15-30 min to 5 min).

Conclusions:

• Guided Correction Software enables efficient labeling of training data for deep learning.
• This approach enhances the performance and generalization capabilities of EC segmentation models.
• The software facilitates improved analysis of diverse post-keratoplasty EC image datasets.