Seeing Through the Mycelium: How Computer Vision Is Unlocking Fungal Secrets

Too long to read? Go for the highlights below.

• Researchers have created the first large-scale dataset for training AI to analyse mycelium growth patterns, comprising over 20,000 high-resolution images of four fungal species

• The technology enables precise measurement of hyphal networks that was previously impossible, opening new avenues for strain selection and environmental monitoring in industrial mycology

• Current AI models struggle with the delicate edges of fungal structures, achieving only 28% accuracy on boundary detection; a challenge that must be solved for real-world applications

---

The study of fungi has long relied on the human eye and microscope. But as mycelium-based materials and medicinal fungi gain commercial traction, researchers need faster, more precise ways to understand how these organisms grow. A new dataset called MyceliumSeg, developed by teams from Wuhan University of Technology and Jilin Agricultural University, represents the first systematic effort to teach computers to see what mycologists see.

The Edge Problem

Training artificial intelligence to recognise mycelium isn't straightforward. Unlike identifying a cat in a photograph, distinguishing delicate fungal filaments from their growth medium requires pixel-by-pixel precision. The semi-transparent nature of hyphal edges (those thread-like structures that form the body of a fungus) makes them particularly troublesome for conventional computer vision algorithms.

The research team assembled 20,176 images capturing the complete lifecycle of species including Ganoderma lucidum (reishi) and Pleurotus ostreatus (oyster mushroom). Using a specialised imaging system with dual lighting, they captured growth patterns from initial germination through to full petri dish colonisation. But gathering images proved to be the easier task: annotating them consumed 37 person-days of expert labour.

Oyster Mushroom Source: MedicineNet

This manual annotation process revealed a fundamental challenge: even trained annotators disagreed on precisely where mycelium boundaries should be drawn. The researchers developed a "disagreement quantification method" to identify particularly difficult samples, which then received collaborative review. Such rigorous quality control reflects the inherent difficulty of the task.

Performance Gaps

When mainstream AI models including U-Net, DeepLabv3, and SegFormer were tested on MyceliumSeg, the results proved humbling. Whilst these algorithms achieved respectable overall accuracy, over 84% for general segmentation, their performance on boundary detection told a different story. The best model, SegFormer, managed only 28.6% accuracy when evaluated specifically on edge precision.

This boundary detection gap matters considerably for practical applications. Accurately measuring where new hyphae emerge enables researchers to quantify growth rates, assess strain characteristics, and monitor responses to environmental conditions. Small errors in edge detection cascade into significant inaccuracies for these downstream analyses.

Industrial Implications

The applications extend beyond academic curiosity. Industrial fermentation of medicinal fungi requires real-time monitoring of mycelial health. Materials science companies developing mycelium-based alternatives to leather and packaging need to understand how different strains respond to production conditions. Plant pathologists studying fungal diseases benefit from automated analysis of infection patterns.

The MyceliumSeg dataset provides a foundation for developing more sophisticated algorithms. The 19,609 unlabelled images enable researchers to explore semi-supervised learning approaches, whilst the carefully annotated samples establish benchmarks for measuring progress. The research team has made both data and code publicly available, lowering barriers for computational biologists entering the field.

Samples and Findings Source: Et al. Qianguang Yuan

Yet the 28% boundary accuracy achieved by current models underscores how far the technology must advance before it can reliably assist mycologists. The gap between general recognition and precise edge detection represents a worthwhile challenge for the computer vision community; one with tangible benefits for both fundamental research and commercial applications in the growing fungal biotechnology sector.