Smart Operation Theatre: An AI-based System for Surgical Gauze Counting

In partnership with Singapore General Hospital, this paper presents an improved AI-based system utilizing a single YOLOv5 model trained on 11,000 images to accurately and efficiently track surgical gauzes in real-time, thereby preventing retained surgical items and addressing previous limitations in accuracy, speed, and usability.

The Gist

Imagine you are baking a massive, complicated cake for a wedding. You have a rule: you must never leave a measuring cup, a spoon, or a piece of parchment paper inside the cake. If you do, it could make the guests sick, ruin the cake, and get you sued.

In the real world, this "cake" is a patient, and the "measuring cups" are medical gauzes (soft cloths used to stop bleeding).

This paper describes a project by a student named Saraf Krish and his team at Nanyang Technological University (NTU), working with Singapore General Hospital (SGH). They built a "Smart Butler" for operating rooms to make sure no gauze is ever left inside a patient.

Here is the story of their invention, explained simply:

1. The Problem: The "Gossypiboma" Monster

Sometimes, during surgery, a doctor or nurse forgets to take a piece of gauze out of the patient's body. This is called Gossypiboma.

• The Consequence: It can cause infections, pain, or even death. It also leads to huge lawsuits for the hospital.
• The Old Way: Currently, nurses act like human accountants. They have to count hundreds of gauze pieces by hand before the surgery starts and again after it ends.
  • The Flaw: Humans get tired, distracted, or make mistakes. One study showed that in 88% of cases where a gauze was left behind, the nurses thought they had counted correctly!
  • The Expensive Fix: Some hospitals use special "smart gauzes" with radio chips (RFID), but these cost 8 times more than normal gauze. Most hospitals can't afford them.

2. The Solution: The AI "Eagle Eye"

The team built a system that acts like a super-vigilant security guard that never blinks.

• How it works: They set up two trays in the operating room.
  • Tray A ("In"): Holds clean, unused gauze.
  • Tray B ("Out"): Holds dirty, used gauze.
• The Magic: Cameras watch these trays 24/7. An Artificial Intelligence (AI) brain, powered by a model called YOLOv5 (think of it as a super-fast pattern recognizer), watches the video feed.
• The Count: Every time a nurse picks up a clean gauze, the AI sees it and subtracts from the "In" pile. Every time a used gauze is put on the "Out" tray, the AI adds to that pile.
• The Goal: At the end of the surgery, the number of gauzes taken out must equal the number put in. If the math doesn't add up, the system screams (visually) to stop the surgery until the missing piece is found.

3. How They Taught the AI (The "School" Analogy)

You can't just give a computer a camera and expect it to know what a bloody piece of cloth looks like. You have to teach it.

• The Classroom: The team went into real operating rooms with doctors. They took thousands of photos of gauze in weird situations: stacked on top of each other, covered in blood, in bright lights, or in shadows.
• The Upgrade: In their first attempt, they used two separate "teachers" (one to find hands, one to find gauze). It was slow and clunky.
• The New Teacher: They combined them into one super-intelligent teacher. They fed it 11,000 images (up from 2,800). Now, it's so fast it can "see" 15 times a second (15 FPS), which is fast enough to keep up with a busy surgery.

4. The "Traffic Light" Dashboard

The system doesn't just count; it talks to the doctors.

• Green Light: "Go ahead, move gauze."
• Yellow Light: "I see a hand reaching in."
• Red Light: "Wait! I'm updating the count."
• The Final Check: The screen shows a simple equation: Total In - Total Out = What's Inside the Patient. If the result is zero, everyone is safe to close the patient up.

5. Why This Matters (The "Wallet" and "Heart" Analogy)

• Saving Money: The team estimates that hospitals lose about $6.7 million a year just because of these mistakes (lawsuits, extra surgeries, etc.). This AI system is much cheaper than the "smart RFID gauze" because it uses regular gauze and just adds a camera.
• Saving Lives: It removes the stress of counting from the nurses, letting them focus on the patient. It acts as a safety net that catches human errors before they become tragedies.

6. The "Not Perfect Yet" Part

The team is honest about what still needs work.

• The Stacking Problem: If a nurse grabs a stack of 5 gauzes at once, the AI sometimes gets confused and thinks it's just one big blob. They are working on teaching the AI to see through the "stack."
• The Workflow: Nurses have to be careful to place gauzes on the tray one by one for the AI to count them perfectly. The team hopes to make the system smarter so it can handle messy stacks in the future.

The Bottom Line

This project is like giving the operating room a digital safety net. It uses modern AI to solve an ancient problem: "Did we leave anything behind?" By making the counting automatic, accurate, and cheap, they hope to make surgeries safer for patients and less stressful for the medical teams.

Technical Summary

1. Problem Statement

The paper addresses the critical medical issue of Gossypiboma (Retained Surgical Items or RSI), where medical gauzes are accidentally left inside a patient's body after surgery. This condition leads to severe complications such as abscesses, septic shock, and even death, while also exposing hospitals to significant legal liabilities and financial penalties (estimated at $1.46 million per incident on average).

Current prevention methods face distinct limitations:

• Manual Counting: Relies on nurses tallying hundreds of gauzes, which is time-consuming, diverts resources from patient care, and has a high error rate (studies show 88% of inaccurate counts are deemed correct).
• RFID-Integrated Gauzes: While accurate, these are approximately eight times more expensive than standard gauzes (adding ~$0.50 per unit plus scanner costs), making widespread adoption financially unfeasible for many hospitals.

The research aims to develop a cost-effective, AI-driven solution to automate gauze counting, thereby eliminating human error without the prohibitive costs of RFID technology.

2. Methodology

The project utilizes a computer vision approach based on Deep Learning and Real-time Object Detection, developed in partnership with Singapore General Hospital (SGH).

Data Collection and Preparation

• Collaboration: Data was gathered in actual operating theatres at SGH under the supervision of medical professionals.
• Dataset: The team collected and annotated 11,000 images (an eightfold increase from previous iterations of 2,800 images). The dataset includes diverse scenarios: blood-stained gauzes, overlapping items, varying lighting conditions, and shadows.
• Annotation: Images were labeled using Roboflow to identify both surgical gauzes and human hands.
• Setup: A dual-tray system was established:
  • "In" Tray: Holds unused, clean gauzes.
  • "Out" Tray: Holds used, blood-stained gauzes.

System Architecture & Technologies

• Core Model: The system employs YOLOv5 (You Only Look Once), a deep learning object detection model.
  • Evolution: The project shifted from using two separate models (one for humans, one for gauze) to a single integrated model capable of detecting both simultaneously.
• Hardware:
  • A mobile frame constructed from aluminum profiles with adjustable arms.
  • High-resolution cameras mounted on the frame.
  • Integrated LED lighting to ensure consistent illumination regardless of ambient OT lighting.
• Software Stack:
  • PyTorch: For loading and running the YOLOv5 model.
  • OpenCV: For video capture, frame resizing, and user interface (UI) overlay.
  • CUDA: To leverage NVIDIA GPUs for accelerated inference, ensuring real-time processing.
  • Multithreading: Implemented to process video streams from two cameras simultaneously without latency.

Operational Logic

The system tracks the movement of gauzes between the "In" and "Out" trays.

• Counting Logic:
  • Total In: Cumulative count of gauzes taken from the "In" tray.
  • Total Out: Cumulative count of gauzes placed on the "Out" tray.
  • In Play (Current Count): Calculated as Total In - Total Out.
• Safety Goal: At the end of surgery, In Play must equal 0, indicating no gauze remains inside the patient.
• User Interface: A "Traffic Light" system provides visual feedback:
  • Green: Safe to add/remove gauze.
  • Yellow: Hand detected (system is monitoring interaction).
  • Red: System updating counts (flashes for ~50ms).
• Manual Override: The system includes a feature for doctors/nurses to manually adjust counts if discrepancies arise, enhancing reliability.

3. Key Contributions

1. Integrated Single-Model Architecture: Unlike previous iterations that used separate models for humans and objects, this study successfully trained a unified YOLOv5 model to detect both, significantly improving efficiency.
2. Dataset Expansion: The training dataset was scaled from 2,800 to 11,000 images, covering complex real-world surgical scenarios (overlaps, blood stains, lighting variations).
3. Performance Optimization: The system achieved a 100% increase in frame rate, jumping from 8 FPS to 15 FPS, making it viable for real-time surgical monitoring.
4. Cost-Effective Alternative: The solution offers a high-accuracy alternative to RFID systems, utilizing standard gauzes and off-the-shelf hardware (cameras/laptops) to drastically reduce implementation costs.
5. Human-in-the-Loop Design: The inclusion of manual count adjustment features acknowledges the dynamic nature of surgery, ensuring the system remains a reliable aid rather than a rigid constraint.

4. Results

• Accuracy & Speed: The integrated model demonstrated improved precision and successfully processed video feeds at 15 FPS, allowing for real-time tracking.
• Functional Validation: The system successfully tracks gauze movement between trays, maintaining a running count of "In Play" items.
• User Feedback: Initial trials with SGH staff confirmed the system's ability to detect gauzes accurately. However, feedback highlighted a need for better handling of stacked gauzes (currently, the system requires gauzes to be separated on the tray) and improved detection of overlaps exceeding 50%.
• Future Iterations: The team plans to upgrade to newer models (e.g., YOLOv9) to further increase speed and accuracy, specifically targeting the detection of stacked items.

5. Significance and Impact

• Patient Safety: By virtually eliminating the risk of retained surgical items, the system directly prevents life-threatening complications and reduces patient mortality and morbidity associated with Gossypiboma.
• Economic Impact: The system offers a solution that is significantly cheaper than RFID alternatives. It aims to save hospitals millions in potential legal settlements and regulatory penalties (estimated at $6.72 million annually for SingHealth Group alone due to retained items).
• Operational Efficiency: Automating the counting process frees up nursing staff to focus on direct patient care rather than manual tallying.
• Scalability: The modular hardware design (mobile frame, standard cameras) allows for easy integration into existing operating theatres without major infrastructure changes.

In conclusion, this project represents a significant step forward in AI-assisted healthcare, successfully translating deep learning theory into a practical, cost-effective tool that enhances surgical safety and hospital operational efficiency.