💻 computer science

Lightweight Faster R-CNN with Standalone BiFPN for Accurate Detection and Counting of Tuberculosis Bacilli in Ziehl-Neelsen Smear Microscopy

This study presents a lightweight Faster R-CNN framework enhanced with a standalone BiFPN that achieves high-accuracy detection and counting of tuberculosis bacilli in Ziehl-Neelsen smear microscopy, offering a computationally efficient AI solution to accelerate diagnosis and reduce variability in tuberculosis control.

Original authors: Mahasti Namira, Ridwan Wicaksono, Titik Nuryastuti

Published 2026-08-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mahasti Namira, Ridwan Wicaksono, Titik Nuryastuti

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to find tiny, sneaky clues hidden inside a giant, messy crime scene. In the world of medicine, one of the most important crimes is Tuberculosis (TB), a disease caused by microscopic bacteria called Mycobacterium tuberculosis. To catch these culprits, doctors use a special "stain" called Ziehl–Neelsen, which turns the bacteria a bright red color against a blue background, kind of like highlighting a suspect in a crowd. But here's the problem: these bacteria are incredibly small, they often huddle together in tight groups, and there are thousands of them on a single slide. A human doctor has to stare at these slides under a powerful microscope for hours, counting every single red speck. It's exhausting work, and when people get tired, they might miss a clue or count the wrong thing.

This is where computer science steps in with a tool called "Deep Learning." Think of Deep Learning as teaching a computer to see and learn, just like a child learns to recognize a cat by looking at many pictures. Specifically, this paper uses a type of computer vision called "Object Detection," which is like giving the computer a magnifying glass and a checklist to find and count specific items in an image. The big question scientists have been asking is: Can we build a computer brain that is smart enough to find these tiny bacteria accurately, but also small and fast enough to run on regular computers in hospitals that don't have super-expensive supercomputers?

The Paper's Story: A Lightweight Detective with a Super-Brain

In this research, a team of scientists from Universitas Gadjah Mada in Indonesia decided to build a new kind of "digital detective" to solve the TB counting problem. Their goal was to create a system that is not only accurate but also "lightweight," meaning it doesn't need a massive, expensive computer to run. They wanted something that could work on standard lab equipment, making it easier for doctors in busy clinics to get help.

The team started with a famous computer vision recipe called Faster R-CNN. You can think of this as a two-step detective process: first, the computer scans the image to find "regions of interest" (places where a clue might be hiding), and second, it zooms in to decide exactly what that clue is. However, the standard version of this detective is heavy and slow, like a detective carrying a giant, heavy backpack full of tools. To fix this, the researchers swapped out the heavy backpack for a MobileNetV3 backbone. Imagine this as replacing the heavy backpack with a sleek, high-tech utility belt. It's much lighter and faster, but the worry was: "Will it still be smart enough to find the tiny bacteria?"

To make sure the lightweight detective didn't miss anything, the team added a special new gadget called Standalone BiFPN (Bidirectional Feature Pyramid Network). Here is a fun way to understand what this does: Imagine the bacteria are like different-sized toys scattered on the floor. Some are tiny, some are big, and some are stuck together. A normal camera might only see the big toys well, or only the small ones. The BiFPN is like a magical lens that lets the computer look at the image from many different "zoom levels" at the same time. It mixes the "big picture" view (to understand the context) with the "super-close-up" view (to see the tiny details). Crucially, it does this in a two-way street, passing information up and down between the zoom levels so the computer never gets confused.

What they found:
The team tested their new "Lightweight Faster R-CNN with Standalone BiFPN" against several other methods, including the heavy, slow version and some other popular fast detectors. The results were impressive. Their new model achieved a detection accuracy (called mAP@0.5) of 93.71%. This means it was incredibly good at finding the bacteria.

But accuracy isn't everything; the computer also needs to count correctly. The researchers compared the computer's count against the count made by expert human technicians. The results showed a near-perfect match, with a correlation score of 0.997. To put that in perspective, if the human expert counted 100 bacteria, the computer almost always counted very close to 100. The average difference (error) was only 0.22 bacteria per image, which is practically zero.

What they ruled out:
The paper explicitly tested a few other ideas to see if they would work better, and it ruled some of them out.

Adding too many gadgets: They tried adding another tool called CBAM (an attention mechanism that helps the computer focus on important parts) along with the BiFPN. Surprisingly, this made things worse. Instead of helping, the computer got "over-sensitive" and started seeing bacteria where there were none (false positives). The paper found that the BiFPN alone was the perfect balance; adding the extra gadget just made the detective too jumpy.
One-step detectors: They also tested a popular fast detector called YOLOv9c. While it was fast, it struggled significantly with the tiny, clustered bacteria, often missing them or getting confused by the noise. The paper suggests that for this specific, tiny-scale job, the two-step "Faster R-CNN" approach is superior to the one-step approach.

How sure are they?
The authors are very confident in their findings based on the data they collected. They didn't just guess; they trained the model on a dataset of 1,265 images (which they expanded to 6,375 using data augmentation techniques like flipping and rotating to make the training tougher). They tested the model against a "gold standard" of manual counts performed by certified medical lab technicians. The statistical tests showed that the agreement between the computer and the humans was "excellent," with a very narrow margin of error (Limits of Agreement of -0.63 to +0.67).

The Bottom Line:
The paper concludes that you don't need a giant, heavy computer to find TB bacteria. By using a smart, lightweight backbone (MobileNetV3) and a specialized feature-fusion tool (BiFPN), they created a system that is both fast (taking only 16.38 milliseconds to analyze an image) and incredibly accurate. This means that in the future, hospitals with limited resources could use this software to help doctors count bacteria quickly and reliably, potentially saving time and lives by catching TB earlier. The researchers note, however, that while the results are strong, the next step would be to test this in real-world clinics across different locations to make sure it works everywhere.

Technical Summary: Lightweight Faster R-CNN with Standalone BiFPN for Tuberculosis Bacilli Detection

Problem Statement
Tuberculosis (TB) remains a critical global health burden, with Ziehl–Neelsen (ZN) stained sputum smear microscopy serving as the primary diagnostic tool in low- and middle-income countries. However, manual microscopy is labor-intensive, prone to inter-observer variability, and susceptible to technician fatigue. While deep learning offers a path to automation, existing solutions face a translational gap: high-performing detectors (e.g., those using ResNet backbones) are computationally prohibitive for resource-constrained laboratory hardware, while lightweight or one-stage detectors often struggle with the extreme foreground-background imbalance and the small, clustered nature of Mycobacterium tuberculosis bacilli. Furthermore, many studies prioritize detection metrics (mAP) over clinical validity, failing to demonstrate statistical agreement with the manual counting protocols used for actual diagnosis.

Methodology
The authors propose a computationally efficient deep learning framework integrating a Bidirectional Feature Pyramid Network (BiFPN) into a Faster R-CNN pipeline, specifically tailored for microscopic bacilli detection.

Architecture: The core model utilizes MobileNetV3-Large as the backbone to minimize parameter count and computational cost while retaining representational capacity for fine structures. Instead of a standard Feature Pyramid Network (FPN), the authors employ a standalone BiFPN as the sole feature fusion module. This design facilitates bidirectional cross-scale information exchange (top-down and bottom-up) with weighted fusion, allowing the network to adaptively prioritize feature scales based on image characteristics (e.g., favoring fine resolution for isolated bacilli and coarser representations for dense clusters).
Data Processing: The study utilized a Kaggle dataset of 1,265 ZN-stained images, expanded to 6,375 via augmentation (rotation, flipping, brightness/contrast adjustments). Preprocessing included RGB conversion, median filtering, and LAB color space enhancement using CLAHE to improve local contrast without altering stain colors.
Annotation Protocol: Ground truth was established through a rigorous verification process involving certified Medical Laboratory Technologists (MLTs). Annotations followed WHO and IUATLD guidelines, ensuring bounding boxes tightly enclosed bacilli while excluding artifacts. For the final clinical validation, a separate dataset was manually annotated by an MLT to serve as the gold standard.
Post-Processing: The detection pipeline applies a confidence threshold of 0.5 and Non-Maximum Suppression (NMS) with a high IoU threshold (0.8) to handle overlapping clusters and prevent double-counting.
Experimental Design: An ablation study compared six variants: Faster R-CNN with ResNet-50, Faster R-CNN with MobileNetV3, MobileNetV3 + CBAM, MobileNetV3 + BiFPN (proposed), MobileNetV3 + CBAM + BiFPN, and YOLOv9c as a benchmark.

Key Contributions

Architectural Innovation: The integration of a standalone BiFPN with a lightweight MobileNetV3 backbone is shown to be more effective for small-object detection than adding attention mechanisms (CBAM) or using heavier backbones. The study demonstrates that optimizing multi-scale feature fusion is more impactful for this task than simply increasing network depth.
Clinical-Centric Validation: Unlike many prior works that rely solely on detection metrics, this study rigorously validates the model against expert manual counts using statistical correlation (Pearson's r, Spearman's $\rho$ ), error metrics (MAE, RMSE), and Bland–Altman analysis (Limits of Agreement).
Efficiency vs. Accuracy Trade-off: The proposed model achieves a balance suitable for edge deployment, reducing the parameter count by over 92% compared to a ResNet-50 baseline while simultaneously improving detection accuracy.

Results

Detection Performance: The proposed MobileNetV3 + BiFPN model achieved the highest performance across all metrics, with 93.71% mAP@0.5, 98.26% F1-Score, 96.64% Precision, and 98.78% Recall.
- Notably, the combination of CBAM and BiFPN resulted in a significant drop in mAP (54.20%) despite high recall, indicating that the attention mechanism caused over-sensitization to artifacts (false positives).
- The ResNet-50 baseline achieved 90.75% mAP but required significantly more computational resources.
- YOLOv9c underperformed (75.54% mAP), struggling with fine-grained microscopic objects and dense clusters.
Computational Efficiency: The proposed model contains only 12.6 million parameters and has a file size of 48.1 MB, representing a 92.3% reduction in parameters and 92.4% reduction in size compared to the ResNet-50 baseline. Inference time was 16.38 ms/image, offering a favorable speed-accuracy balance.
Clinical Correlation: The proposed model demonstrated near-perfect agreement with expert manual counts:
- Pearson's r: 0.997
- MAE: 0.22
- RMSE: 0.47
- $R^2$ : 0.99
- Bland–Altman Analysis: The model showed minimal bias (0.022) and the narrowest Limits of Agreement (±0.67), confirming statistical and clinical interchangeability with manual counting. In contrast, the ResNet-50 model showed a much wider LOA range (±8.32) and negative $R^2$ , indicating poor correlation with clinical grading despite decent detection scores.

Significance and Claims
The paper claims that its primary contribution lies in bridging the gap between algorithmic performance and clinical utility. By demonstrating that a lightweight architecture with optimized feature fusion can outperform heavy baselines in both detection accuracy and statistical agreement with manual counting, the study offers a viable solution for real-world deployment in resource-limited settings. The authors assert that their framework provides "reliable diagnostic cues" that align with routine laboratory reporting needs, potentially accelerating case detection and reducing diagnostic variability. The work emphasizes that for microscopic bacilli detection, careful optimization of multi-scale feature representation is more critical than network depth, enabling the creation of scalable, effective AI assistance tools for TB control.

The Paper's Story: A Lightweight Detective with a Super-Brain

Technical Summary: Lightweight Faster R-CNN with Standalone BiFPN for Tuberculosis Bacilli Detection

More like this