Automated Classification and Quantification of Fungal Spores in Aerobiological Sampling Using Deep Learning

Problem Statement

Routine aerobiological monitoring of fungal spores is critical for environmental surveillance and public health, particularly given the link between airborne spores and respiratory conditions like asthma and allergic rhinitis. However, current assessment methods rely on manual microscopic counting, a process that is time-consuming, prone to inter-observer variability, and results in delayed data availability. While previous automated attempts have utilized classical computer vision or focused on single-genus detection (e.g., Alternaria), these approaches often struggle with particle overlap, require manual feature engineering, and lack the robustness to handle the high taxonomic diversity found in real-world atmospheric samples. There is a distinct need for a scalable, multi-genus deep learning framework capable of automating detection and quantification across diverse fungal taxa.

Methodology

The study developed and validated a deep learning-based workflow for the automatic detection and counting of eight fungal genera (Cladosporium, Coprinus, Curvularia, Didymella, Ganoderma, Leptosphaeria, Nigrospora, and Xylaria) using microscopic images from Hirst-type volumetric sampling in Santa Marta, Colombia.

1. Dataset Construction and Preprocessing:

• Source: 325 original images (400× magnification) were subdivided into 1,281 quadrants (1296 × 972 pixels).
• Annotation: Performed using Roboflow, with bounding boxes assigned to 2,252 spore instances across the eight classes.
• Imbalance Mitigation: The original dataset exhibited class imbalance (e.g., Coprinus at 20.1% vs. Nigrospora at 9.1%). To address this, the authors implemented:
  • Stratified Splitting: Images were re-split into 70% training, 20% validation, and 10% testing based on the predominant spore type.
  • Offline Augmentation: Synthetic samples were generated for underrepresented classes using the Albumentations library (geometric, photometric, and degradation transformations) to target ~320 instances per class.
  • Online Augmentation: During training, Mosaic (to increase object density) and Mixup (for regularization) were applied.
• Final Dataset: The training set was expanded to 1,574 images (2,963 instances), while validation and test sets remained composed of original, un-augmented images to ensure unbiased evaluation.

2. Model Architectures:
Three detection strategies were compared:

• YOLO26: Evaluated in Medium (YOLO26m) and Large (YOLO26l) variants. Input resolution was 1280 × 1280 pixels to preserve morphological details.
• ViT-B/16 + Faster R-CNN: A hybrid approach using a Vision Transformer for feature extraction (with the first six layers frozen) coupled with Faster R-CNN for region proposal and classification.
• AlexNet + Faster R-CNN: A modified AlexNet with batch normalization coupled with Faster R-CNN, using a 640 × 640 pixel input resolution.

3. Training and Evaluation:
Models were trained with specific hyperparameters (e.g., YOLO variants used MuSGD, ViT/AlexNet used AdamW/SGD). Performance was evaluated using Precision, Recall, F1-score, and mean Average Precision (mAP@0.5 and mAP@0.5:0.95). Given the application context, Recall and Precision were prioritized over strict localization metrics.

Key Results

1. Model Performance Comparison:

• YOLO26l emerged as the superior model, achieving the highest Recall (0.812), Precision (0.823), mAP@0.5 (0.846), and F1-score (0.817) on the test set.
• YOLO26m followed closely with slightly lower metrics.
• ViT-B/16 + Faster R-CNN showed intermediate performance but suffered from a significant drop in precision (0.502).
• AlexNet + Faster R-CNN yielded the lowest performance across all metrics (Recall 0.548, Precision 0.377).
• Genus-Specific Performance: YOLO26l demonstrated robustness for Curvularia, Didymella, Ganoderma, Nigrospora, and Xylaria. Cladosporium remained the most challenging class, with the lowest recall (0.429 in testing) and highest rate of false negatives (54% of instances missed).

2. Operational Validation (Automated vs. Manual):

• An external set of 516 unseen images was analyzed. YOLO26l maintained a precision of 0.803 and a recall of 0.659.
• Correlation: Automated and manual counts showed positive and significant correlations for all genera (Spearman's ρ = 0.387–0.852; p < 0.001). The correlation was strongest for Curvularia (0.852) and weakest for Cladosporium (0.387).
• Efficiency: The automated system processed 516 images in 1 minute 39 seconds (0.19 s/image), compared to 2.8 hours (19.5 s/image) required for manual counting. This represents a 102-fold increase in processing speed.

3. Error Analysis:
The primary source of error was false negatives (missed detections) rather than misclassification between genera. The model generally assigned the correct genus once a spore was detected. The difficulty in detecting Cladosporium was attributed to its high morphological variability and visual similarity to background structures.

Key Contributions

• Multi-Genus Framework: The study moves beyond single-genus detection to validate a system capable of simultaneously identifying eight distinct fungal genera, addressing the taxonomic diversity inherent in aerobiological sampling.
• Dataset Curation: Creation and augmentation of a specialized dataset for fungal spores, including strategies to handle class imbalance and low object density in microscopic fields.
• Architectural Benchmarking: A comparative analysis demonstrating that single-stage detectors (YOLO26) outperform hybrid and region-proposal-based models (Faster R-CNN variants) for this specific task, particularly when balancing computational efficiency and detection accuracy.
• Operational Scalability: Demonstration of a workflow that reduces analysis time by two orders of magnitude, facilitating the potential for real-time or near-real-time bioaerosol surveillance.

Significance and Claims

The authors claim that this work provides a reference framework for scaling aerobiological monitoring. By automating the most labor-intensive stage of the process (microscopic counting), the system addresses the bottleneck that limits the geographic and temporal coverage of current monitoring programs.

The paper modestly asserts that while performance varies by genus, the system successfully reproduces general abundance patterns observed in manual counts. It positions deep learning not merely as a classification tool but as an operational instrument for environmental surveillance and public health risk assessment. The authors acknowledge that further improvements are needed for difficult classes like Cladosporium but conclude that the proposed approach significantly strengthens the capacity for routine, timely, and reproducible fungal spore monitoring.