Abstract
PURPOSE: Corneal confocal microscopy (CCM) is a powerful tool for detecting early signs of neurodegenerative diseases by analyzing the corneal nerve fiber morphology. Current automated analysis tools, such as ACCMetrics, are outdated and lack extensibility. Most deep learning-based models are limited to single tasks and are rarely open source. This paper proposes SuperCCM, a fully automated, modular, open source Python package, for the comprehensive analysis of CCM images.
METHODS: SuperCCM integrates a complete analysis pipeline comprising image segmentation, skeletonization, topological modeling, and quantitative parameter extraction. The system provides five default modules and supports the easy integration of custom algorithms. A finely annotated dataset (SuperCCM-FineSet, 210 images from 34 participants) was developed to train and evaluate the segmentation models using multistage training with coarsely labeled data from the CORN-1 dataset. Model performance was assessed using clDice, and the morphological parameters were compared with manual annotations.
RESULTS: SuperCCM achieved high segmentation accuracy on an independent test set, with the best encoder-decoder combination (VGG-11 + U-Net) reaching a clDice score of 0.879. For morphological quantification, SuperCCM demonstrated strong consistency with manual annotations, yielding higher intraclass correlation coefficients and lower errors across most parameters compared with ACCMetrics.
CONCLUSIONS: SuperCCM offers an extensible, open source, clinically relevant framework for CCM image analysis. Bridging algorithm development and clinical research enable the accurate and automated quantification of corneal nerve parameters and support the integration of novel deep learning models.
TRANSLATIONAL RELEVANCE: SuperCCM promotes reproducible research and supports the development of new diagnostic tools and deep learning models for corneal confocal microscopy imaging.