Object Detection Techniques for Live Monitoring of Amoeba in Phase-Contrast Microscopic Images

This study develops and evaluates nine Detectron2 and six YOLO v10 deep-learning models for real-time, single-class detection of amoebae in phase-contrast microscopic images, utilizing a diverse dataset of 88 images and 4,131 annotations to balance speed and accuracy for automated bio-imaging analysis.

The Gist

Imagine you are a scientist trying to watch tiny, single-celled creatures called amoebas move around under a microscope. These little guys are like the "immune system scouts" of the microscopic world. To study them, you need to take photos and count them, track where they go, and see how they behave.

In the old days, a human would have to sit there for hours, squinting at a screen, drawing boxes around every single amoeba. It's boring, slow, and humans get tired and make mistakes. Plus, to see them clearly, you often have to shine very bright lights on them, which can actually hurt or scare the amoebas, changing their behavior.

This paper is about teaching a computer to do this job instead of a human, and doing it so fast and gently that the amoebas don't even notice.

The Problem: The "Halo" Effect

The photos taken with these special microscopes (called phase-contrast) are tricky. The amoebas look almost exactly like the background, and they have a weird glowing "halo" around them, like a ghostly ring. It's like trying to find a clear glass marble sitting on a glass table; they blend in perfectly.

The Solution: Two Teams of Digital Detectives

The researchers set up a competition between two famous teams of "digital detectives" (AI models) to see which one is best at finding these ghosts:

1. Team "Slow and Steady" (Detectron2 / Faster R-CNN):

   • How they work: Imagine a detective who looks at a photo, points to a spot, says "Hmm, that looks interesting," and then zooms in to take a second, very careful look before saying, "Yes, that's definitely an amoeba."
   • Pros: They are very accurate. They rarely miss an amoeba or confuse a yeast cell for an amoeba.
   • Cons: They take a little longer to think.

2. Team "Fast and Furious" (YOLOv10):

   • How they work: This team is like a security guard scanning a crowd. They look at the whole picture in one split second and shout, "There's an amoeba there! And one there! And one there!"
   • Pros: They are incredibly fast. If you need to track amoebas in real-time (like watching a live video), they are great.
   • Cons: Sometimes they get a little excited and count the same amoeba twice, or they might mistake a speck of dust for an amoeba.

The Experiment

The researchers fed these two teams a "training school" consisting of 88 photos with over 4,000 amoebas marked by humans. They tested the teams on different sizes of photos and different lighting conditions.

Here is what they found:

• Accuracy: The "Slow and Steady" team (Detectron2) won the accuracy contest. They were slightly better at distinguishing a real amoeba from background noise or yeast cells.
• Speed: The "Fast and Furious" team (YOLO) was much quicker. If you needed to process a video stream instantly, YOLO was the winner.
• The "Double Counting" Issue: The fast team sometimes saw the same amoeba multiple times (like seeing a reflection in a mirror and thinking it's a second person). The slow team was better at realizing, "No, that's just the same guy."

The Big Takeaway

The researchers concluded that both teams are good, but they serve different purposes:

• If you are doing a live experiment where you need to watch amoebas move in real-time without lag, use the Fast Team (YOLO). It's fast enough to keep up, even if it's not perfect.
• If you are doing detailed research where you need to know exactly how many amoebas there are and what they look like, use the Slow Team (Detectron2). It's more careful and precise.

Why This Matters

By using these AI tools, scientists can:

1. Save Time: No more hours of manual counting.
2. Save the Amoebas: Because the AI is so good at spotting them in dim light, scientists don't need to blast them with bright, harmful light. The amoebas stay happy and act naturally.
3. Understand Disease: Since these amoebas act like human immune cells, understanding them better helps us understand how our bodies fight infections.

In a nutshell: The paper is a guidebook for scientists on how to choose the right "robot assistant" to watch tiny cells, balancing the need for speed against the need for perfect accuracy.

Technical Summary

1. Problem Statement

The research addresses the challenge of automating the detection and live tracking of Dictyostelium discoideum (amoeba) in phase-contrast microscopic images.

• Context: Manual annotation is time-consuming, prone to human error, and limits the scalability of biological research. Furthermore, reliance on high-intensity fluorescence for labeling can adversely affect cell activity.
• Specific Challenges:
  • Low Contrast: Phase-contrast imaging produces "halo" artifacts and "shade-off" effects, resulting in low contrast between the amoeba and the background.
  • Inhomogeneity: Images vary significantly due to different acquisition times, setups, and environmental conditions.
  • Dense Populations: Amoebas often overlap or touch, making segmentation and individual tracking difficult.
  • Data Scarcity: The study utilizes a relatively small dataset (88 images, 4,131 annotations), necessitating robust models that can generalize without massive data volumes.

2. Methodology

Dataset and Preprocessing

• Data Source: 88 phase-contrast images of D. discoideum captured using a Nikon Ti-e microscope (60× objective) and a Photometrics 95B sCMOS camera.
• Variability: Images vary in resolution (1024–2048 pixels) and include artifacts like yeast stacks and circular fields of view.
• Annotation: Manual polynomial annotation was performed using the "Make Sense" platform for a single class ("amoebas"). Annotations were exported in COCO JSON format (for Detectron2) and YOLO format.
• Split: 64 images (3,156 annotations) for training; 24 images (975 annotations) for validation.
• Preprocessing: Images were rescaled to a maximum dimension of 1024 pixels to balance GPU memory constraints while maintaining aspect ratios.

Model Architectures

The study compared two primary frameworks: Detectron2 (Facebook AI Research) and Ultralytics YOLO (v10).

1. Detectron2 (Two-Stage Detectors):

   • Faster R-CNN: Tested with 7 variants using ResNet-50 and ResNet-101 backbones combined with different feature extraction heads:
     • C4 (Conv4 backbone)
     • DC5 (Dilated Conv5)
     • FPN (Feature Pyramid Network)
   • RetinaNet: Tested with ResNet-50 and ResNet-101 backbones.
   • Training: All models used a 3× learning rate schedule (300k iterations) and pre-trained weights from the COCO dataset.

2. YOLOv10 (One-Stage Detector):

   • Tested 6 variants ranging from Nano (n) to Extra-Large (x): yolov10n, s, m, b, l, x.
   • Utilized the latest YOLOv10 architecture, known for improved efficiency over previous versions.

Evaluation Metrics

• Intersection over Union (IoU): Measured overlap between ground truth and predicted boxes (threshold set at 0.50).
• Average Precision (AP@0.50): The primary metric for accuracy, summarizing the Precision-Recall curve at an IoU of 0.50.
• Inference Speed: Measured in seconds per image for two resolutions (4128px and 1200px).
• Hardware: Training and inference were performed on a PC with an Intel i7-13650HX CPU and NVIDIA GeForce RTX 4060 GPU.

3. Key Contributions

• Comprehensive Benchmarking: A direct comparison of state-of-the-art two-stage (Faster R-CNN, RetinaNet) and one-stage (YOLOv10) detectors specifically tailored for the difficult conditions of phase-contrast amoeba imaging.
• Architecture Analysis: Evaluation of how backbone depth (ResNet-50 vs. 101) and feature extraction strategies (C4, DC5, FPN) impact performance on small, heterogeneous biological datasets.
• Identification of Over-segmentation Issues: The study highlights a specific failure mode where YOLO models tend to over-segment or detect single amoebas multiple times, complicating live tracking, whereas Detectron2 models provided more distinct object localization.
• Practical Guidelines: Provides specific recommendations for researchers balancing the trade-off between real-time inference speed and detection accuracy in low-contrast microscopy.

4. Results

Accuracy (AP@0.50)

• Detectron2 Superiority: Generally, Detectron2 models outperformed YOLOv10 in accuracy.
  • Best Performer: Faster R-CNN + R50-DC5 achieved the highest accuracy at 89.41%.
  • Runner-up: Faster R-CNN + R50-FPN (88.75%) and Faster R-CNN + R101-DC5 (88.78%).
  • YOLO Performance: The best YOLOv10 variant was yolov10m with 88.03% (Note: The text contains a discrepancy in the abstract vs. results section regarding the exact YOLO score, but consistently states YOLO was slightly lower than the best Detectron2 models).
• Backbone Depth: Contrary to the expectation that deeper networks (ResNet-101) always yield better results, the study found that R50 backbones often matched or exceeded R101 in speed with comparable accuracy, suggesting diminishing returns for deeper networks on this specific dataset size.

Speed and Efficiency

• Inference Time: YOLOv10 models were generally faster in inference than Faster R-CNN.
  • YOLOv10-n was the fastest (0.27s on 1200px images).
  • Faster R-CNN + R50-FPN was the fastest Detectron2 model (0.41s on 1200px images).
• Training Dynamics: YOLO models reached early stopping criteria faster than Detectron2 models, which utilized the full 300k iterations.

Qualitative Observations

• Tracking Suitability: Detectron2 models were found to be more suitable for live tracking because they detected individual amoebas distinctly. YOLO models frequently produced multiple overlapping bounding boxes for a single amoeba, requiring complex post-processing to resolve.
• Robustness: Both frameworks handled inhomogeneity at the field-of-view edges well.

5. Significance and Conclusion

• Biomedical Impact: The study demonstrates that automated deep learning can effectively replace manual annotation for amoeba monitoring, potentially reducing the need for high-intensity fluorescence labeling and enabling long-term live cell studies.
• Model Selection Strategy:
  • For Maximum Accuracy: Use Faster R-CNN with a DC5 backbone (specifically R50-DC5).
  • For Real-Time/Speed: Use YOLOv10-s or Faster R-CNN with FPN if a slight drop in accuracy is acceptable for faster inference.
• Future Outlook: While the current results are promising, the authors note that performance could be further improved with larger datasets and that the choice of framework should depend on whether the priority is pixel-level precision (Detectron2) or raw inference speed (YOLO).

In summary, this paper provides a critical evaluation of object detection tools for phase-contrast microscopy, concluding that while YOLO offers speed, Detectron2-based Faster R-CNN models currently offer superior accuracy and reliability for the specific challenges of tracking overlapping, low-contrast biological cells.