End-to-end plaque counting and virus titration from laboratory plate images with deep learning

Technical Summary: End-to-end plaque counting and virus titration from laboratory plate images with deep learning

Problem Statement
Plaque assays remain the gold standard for measuring virus infectivity; however, the quantification of plaques from plate images is traditionally performed via visual inspection. This manual process is labor-intensive, subjective, prone to inter-operator variability, and difficult to standardize across different analysts and laboratories. While previous software solutions exist (e.g., Plaque2.0, Viridot, Fiji/ImageJ macros), they often rely on controlled imaging conditions, specialized optics (such as fluorescence or holography), or laborious parameter tuning. These limitations restrict their adoption for routine, label-free cytopathic effect (CPE) plaque assays captured as simple photographs.

Methodology
The authors propose an end-to-end, computer-aided workflow for virus titration directly from laboratory plaque assay images. The pipeline integrates two deep learning models derived from the Segment Anything Model (SAM) family, designed to operate under heterogeneous imaging conditions without specialized hardware.

1. Well Segmentation (SAM2-based):

   • Goal: To automatically detect and localize individual assay wells within a full-plate image, regardless of plate format (6-well or 12-well) or lighting conditions.
   • Model: Utilizes the SAM2 mask generator (specifically the sam2.1_hiera_base_plus checkpoint) in a zero-shot setting.
   • Process: Input images are rescaled, and candidate masks are generated. A custom filtering stage applies geometric constraints (area > 100 pixels, circularity > 0.7) to remove spurious detections. Validated masks are mapped to the original resolution to determine the grid layout (row/column), enabling association with experimental dilution schemes.
   • Training: No fine-tuning was applied; the model relies on pre-trained weights.

2. Plaque Segmentation (SAM-based):

   • Goal: To detect and enumerate plaques within each cropped well.
   • Model: Uses the SAM image encoder (ViT-B backbone) as a frozen feature extractor, paired with a lightweight, task-specific convolutional decoder.
   • Training: Only the decoder parameters were optimized using PyTorch Lightning, Adam optimizer, and binary cross-entropy loss. The model was trained on a mixed dataset of Mayaro virus (MAYV), Coxsackievirus B3 (CVB3), and Vaccinia virus (VACV) images. Extensive data augmentation (geometric and color-based) was employed to improve generalization.
   • Post-processing: Since the model outputs probability maps without explicit instance separation, a dedicated post-processing pipeline is applied. This includes watershed segmentation on distance transforms, Hough circle detection for geometric validation, circularity analysis, DBSCAN clustering to merge nearby peaks, and noise filtering (discarding detections < 10×10 pixels). This step specifically addresses the challenge of separating overlapping plaques.

3. Titration Calculation:

   • The final plaque counts per well are combined with user-provided dilution and inoculation volume data to automatically compute Plaque-Forming Units per milliliter (PFU/mL).

Key Contributions

• Dual SAM-Guided Architecture: A novel workflow combining SAM2 for robust well localization and a fine-tuned SAM variant for plaque segmentation, reducing the annotation burden while maintaining accuracy.
• Integrated Web Platform (Titra): The pipeline is integrated into a web-based platform (Titra) that allows for automated ML-based analysis, rapid quality control, and experiment management.
• Curated Dataset (MAYV/CVB3): The creation and planned public release of a dataset comprising plaque assay images of MAYV and CVB3, captured under routine laboratory conditions using smartphone photography on 6- and 12-well plates.
• Open-Source Release: The system is designed to be open-sourced upon manuscript acceptance to facilitate reproducible and scalable analysis.

Results
The workflow was evaluated on a mixed dataset including private MAYV/CVB3 images and public VACV images (VACVPlaque dataset).

• Well Detection: The model demonstrated high precision and recall across both datasets. For MAYV/CVB3, recall was 0.9898 and precision 0.9614; for VACV, recall was 0.9920 and precision 0.9841 (at IoU threshold 0.5). Failure cases were primarily limited to wells that were almost completely transparent with minimal staining.
• Plaque Detection:
  • Correlation: Strong linear correlations were observed between automated and manual counts. For MAYV/CVB3, the Pearson correlation coefficient was 0.98 (N=180). For VACV, it was 0.88 (N=126), with lower performance attributed to high-density wells where plaque overlap is extensive.
  • Inter-Expert Variability: When compared to four independent experts, the automated method showed high concordance (Pearson r ranging from 0.909 to 0.999). In low-density wells, discrepancies were often comparable to the variability observed between human experts. In high-density wells, the model tended to undercount due to confluence, a limitation shared by human annotators.
  • Error Analysis: Symmetric relative plaque-count errors were generally low (< 0.2). The false-positive rate in empty wells was higher for MAYV/CVB3 (0.74) than for VACV (0.50).
• PFU/mL Estimation: For the MAYV/CVB3 dataset, automated PFU/mL values showed strong agreement with manual calculations (Pearson r = 0.975). Bland-Altman analysis showed a mean difference of 0.06 with 95% limits of agreement ranging from -0.31 to 0.44.
• Performance: Without GPU acceleration, well detection takes approximately 4 seconds, and plaque segmentation takes roughly 4 seconds per well.

Significance and Claims
The paper claims that the proposed system provides a practical, scalable, and reproducible tool for laboratory plaque assay analysis. By leveraging foundation models, the workflow achieves virus-agnostic quantification across different viral species (MAYV, CVB3, VACV) and plate formats without requiring specialized imaging hardware.

The authors emphasize that the system substantially reduces manual annotation effort and inter-operator variability. While acknowledging limitations in extremely high-density (confluent) wells, the study demonstrates that the automated approach yields results highly concordant with expert manual counting and standard titration calculations. The release of the open-source tool and the curated dataset aims to enable audit-ready, reproducible viral quantification for the broader research community.