A Mycelium Dataset with Edge-Precise Annotation for Semantic Segmentation.

Area of Science:

• Computer Vision
• Mycology
• Image Segmentation

Background:

• Accurate mycelium segmentation in petri dish images is crucial for mycology research, enabling quantification of growth patterns and strain characteristics.
• Current computer vision applications are limited by scarce public datasets and the complexity of segmenting low-contrast mycelium edges.

Purpose of the Study:

• To establish the first large-scale benchmark dataset, MyceliumSeg, for mycelium segmentation in mycology.
• To overcome data scarcity and annotation challenges in mycelium image analysis.

Main Methods:

• Created MyceliumSeg dataset with 20,176 diverse images of four fungal species across their growth cycles.
• Developed a novel annotation framework with refined guidelines and disagreement resolution for 567 pixel-level labeled samples.
• Benchmarked mainstream deep learning models using classic and boundary-aware segmentation metrics.

Main Results:

• MyceliumSeg provides a comprehensive resource for developing and evaluating mycelium-specific segmentation algorithms.
• The dataset facilitates research into strain features, environmental adaptability, and responses to stimuli.
• Evaluated deep learning models demonstrate the dataset's utility for advancing segmentation techniques.

Conclusions:

• MyceliumSeg is a valuable resource addressing critical bottlenecks in computer vision for mycology.
• The dataset will accelerate the development of accurate mycelium segmentation tools.
• Facilitates deeper understanding of fungal growth and behavior through advanced image analysis.