Progressive Distillation for Incremental Learning in Corneal Confocal Microscopy Segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Corneal confocal microscopy (CCM) captures morphological changes linked to ocular and systemic diseases.
• Current CCM segmentation methods are often single-task, limiting clinical application.
• A standardized benchmark is lacking for evaluating new CCM analysis techniques.

Purpose of the Study:

• To develop a novel incremental learning approach for multi-structure segmentation in CCM images.
• To introduce a new benchmark dataset and evaluation framework for CCM image analysis.
• To enhance the clinical applicability of CCM by improving segmentation accuracy and standardization.

Main Methods:

• Proposed a data fingerprint distillation (FIND) module for encoding task-relevant knowledge via structural importance mapping.
• Developed a progressive task-guided adapter learning (ProTA) strategy for refining structure representation through staged distillation.
• Created a comprehensive benchmark dataset with manual annotations and an evaluation tool.

Main Results:

• The proposed method achieved state-of-the-art performance in segmenting all corneal structures.
• Demonstrated plug-and-play capability across four other medical image modalities, indicating general utility.
• The new benchmark facilitates unified evaluation of segmentation methods.

Conclusions:

• The novel incremental learning approach significantly improves multi-structure segmentation in CCM images.
• The developed benchmark and tools support standardization and advancement in CCM analysis.
• The method shows potential as a general incremental learning tool for medical image analysis.