Toward Optimal Adenovirus Detection Using YOLO26

Technical Summary: Toward Optimal Adenovirus Detection Using YOLO26

Problem Statement
Transmission electron microscopy (TEM) is a critical tool for characterizing adenovirus particles, particularly for gene therapy vector production. However, accurate automated detection of adenoviruses in TEM images is hindered by visual complexities such as broken particles, debris, and staining artifacts. While deep learning, specifically YOLO-based detectors, has shown promise in general object detection, its application to adenovirus detection remains under-investigated. Furthermore, the specific impact of different data augmentation strategies on the performance of the latest YOLO26 architecture for this specific domain has not been systematically explored.

Methodology
This study conducted a systematic benchmarking experiment to identify the optimal configuration for detecting adenoviruses in TEM images using the YOLO26 model.

• Dataset: The research utilized a subset of the published TEM virus dataset, focusing exclusively on the adenovirus class. The original dataset, containing 1245 images of 22 virus classes, was re-annotated. Since the source data provided only particle center coordinates, a custom Python script generated YOLO-compatible bounding boxes (60 × 60 pixels) centered on these coordinates. The dataset was split into training (40 images), validation (13 images), and test (14 images) sets.
• Model Variants: Five variants of the YOLO26 architecture were evaluated: nano (n), small (s), medium (m), large (l), and extra-large (x). All were trained under identical conditions on an HPC cluster equipped with NVIDIA A100 GPUs.
• Data Augmentation Setups: Four distinct augmentation configurations were defined and compared:
  1. No Augmentation Setup (NAS): Baseline with no transformations.
  2. Geometric Augmentation Setup (GAS): Included rotation, translation, scaling, horizontal/vertical flipping, and shear.
  3. Geometric + Mosaic Augmentation Setup (GMAS): Included all GAS transformations plus Mosaic composition.
  4. Default Augmentation Setup (DAS): The standard augmentation pipeline implemented in YOLO26.
• Evaluation Metrics: Performance was quantified using Precision, Recall, mAP@50, and mAP@50–95. Training convergence, training time, and inference time were also analyzed.

Key Contributions
The paper makes two primary contributions:

1. The design and systematic benchmarking of four specific data augmentation setups (NAS, GAS, GMAS, DAS) across five YOLO26 model size variants.
2. The empirical determination of the most effective configuration for adenovirus detection in TEM images, challenging the assumption that larger models or default settings are inherently superior.

Experimental Results

• Detection Performance: The YOLO26n (nano) variant trained with the GAS configuration achieved the highest overall detection accuracy. It recorded an mAP@50–95 of 0.44, mAP@50 of 0.69, and Precision of 0.86.
• Model Size Impact: Contrary to expectations, increasing the model size did not yield proportional performance improvements. The nano variant generally outperformed the small, medium, large, and extra-large variants across most metrics and augmentation setups.
• Augmentation Impact:
  • GAS proved most effective for the YOLO26n model.
  • GMAS (adding Mosaic) did not consistently improve performance and was associated with less stable convergence and lower qualitative confidence scores.
  • NAS yielded the highest mAP@50–95 for the s, m, and l variants, while DAS provided the highest Recall (0.65) overall.
• Efficiency: Training and inference times correlated strongly with model size, with YOLO26n being the fastest. Interestingly, the GAS and GMAS setups frequently reduced training time compared to DAS and NAS (e.g., YOLO26x training time dropped from 29.31 minutes with DAS to 20.59 minutes with GMAS).
• Convergence: YOLO26n with GAS demonstrated stable convergence curves. In contrast, GMAS exhibited larger fluctuations in validation losses and detection metrics.

Significance and Claims
The authors conclude that optimal adenovirus detection using YOLO26 does not require the largest model size or the most complex augmentation pipelines. Instead, the study identifies a specific combination—YOLO26n with Geometric Augmentation Setup (GAS)—as providing the most favorable balance between detection accuracy, convergence stability, and computational efficiency.

The paper emphasizes that the visual characteristics of adenovirus particles can be effectively learned by a compact model, suggesting that larger models offer limited benefit for this specific dataset. The findings provide practical guidance for researchers aiming to deploy YOLO26 for virus detection in TEM images, advocating for the selection of appropriate model-augmentation pairings over a "bigger is better" approach.

Limitations and Future Work
The authors modestly note that these findings are specific to the single adenovirus TEM dataset used. They do not claim the superiority of GAS and YOLO26n generalizes to other virus types or electron microscopy conditions. Future work is suggested to validate these trends on additional datasets, other object detection architectures, and diverse virus classes.