Technical Report: Automated Optical Inspection of Surgical Instruments

This technical report details a collaboration with industry leaders in Pakistan's Sialkot surgical cluster to develop an Automated Optical Inspection system using deep learning models (YOLOv8, ResNet-152, and EfficientNet-b4) on a new dataset of 4,414 images to detect manufacturing defects in surgical instruments, thereby enhancing patient safety and manufacturing quality.

The Gist

Imagine a world where the tools a surgeon uses to save a life are as reliable as a Swiss watch. Now, imagine that some of those tools have tiny, invisible cracks, rust spots, or scratches that could cause a patient serious harm. This is the reality for the surgical instrument industry, particularly in Pakistan, a global hub for making these life-saving tools.

This technical report is essentially a story about how a team of researchers, working with two major Pakistani companies (Daddy D Pro and Dr. Frigz), built a "super-eyes" system to catch these flaws before they ever reach an operating room.

Here is the breakdown of their project in simple, everyday language:

1. The Problem: The "Human Eye" is Tired

For decades, checking surgical tools (like scalpels, scissors, and forceps) has been done by humans. Imagine a factory worker staring at thousands of shiny metal tools under bright lights, day after day.

• The Issue: Humans get tired. Their eyes get strained. They might miss a tiny scratch or a microscopic pore (a tiny hole caused by rust) because it looks like a reflection.
• The Risk: If a rusty or cracked tool is used in surgery, it can cause infections or break inside a patient. This is a nightmare for both the manufacturer (who loses money) and the patient (who risks their life).

2. The Solution: Teaching a Computer to "See"

The researchers decided to replace the tired human eye with a Digital Super-Eye. They built an Automated Optical Inspection (AOI) system. Think of this as a very strict, never-tiring security guard who looks at every single tool with a high-powered camera and a brain powered by Artificial Intelligence (AI).

3. How They Built the "Brain" (The AI)

To teach the computer what a "bad" tool looks like, they had to do a few things:

• The Classroom (Data Collection): They gathered thousands of photos of surgical tools from their industry partners. They took pictures of tools that were perfect and tools that were broken (with rust, cracks, or scratches).
• The Teacher (Annotation): Experts from the companies looked at these photos and drew boxes around the defects, labeling them: "This is a crack," "This is rust," "This is a pore." This is like a teacher showing a student flashcards and saying, "See this red spot? That's a mistake."
• The Gym (Data Augmentation): To make the AI smarter, they didn't just show it the original photos. They "tricked" the system by rotating the images, changing the brightness, and adding digital noise. This is like training an athlete in the rain, the wind, and the sun so they can perform well in any weather. The AI learned to spot a scratch even if the light was dim or the tool was turned sideways.

4. The Two-Step Detective Process

The AI doesn't just guess; it follows a smart, two-step logic:

1. Step 1: Who are you? First, the AI looks at the tool and says, "Ah, this is a pair of scissors," or "This is a forcep." It knows that a crack on a scissor looks different than a crack on a needle holder.
2. Step 2: What's wrong with you? Once it knows the tool type, it switches to a specialized "detective mode" for that specific tool. It looks for the specific defects that usually happen to that tool.

5. The Results: The "YOLO" Champion

The team tested several different AI models (think of them as different types of detectives). They found that a model called YOLOv8 (which stands for "You Only Look Once") was the clear winner.

• Why? It was incredibly fast and incredibly accurate. It could look at a tool and say, "This is a scissor, and it has a rust spot," in a fraction of a second with near-perfect accuracy (over 99%).
• The Comparison: Other models were either too slow or missed subtle details. YOLOv8 was the perfect balance of speed and precision.

6. The Final Product: "SurgScan"

The researchers didn't just stop at the math; they built a real website called SurgScan.

• How it works: A factory worker or hospital staff member can upload a photo of a tool onto the website.
• The Magic: The system instantly analyzes the image and tells them: "Pass" (it's clean) or "Fail" (it has a defect).
• The Dashboard: Managers can see a big screen showing how many tools were checked, how many were defective, and which batches need to be fixed.

Why Does This Matter?

Think of this project as installing a quality control filter for the global supply chain.

• For Pakistan: It protects their reputation as a top exporter of surgical tools.
• For Hospitals: It ensures that the tools they use are safe, reducing the risk of infections.
• For Patients: It adds an invisible layer of safety, ensuring that the tools saving their lives are in perfect condition.

In short, this paper describes how a team used cameras, AI, and a lot of data to build a tireless, super-accurate inspector that ensures the tools saving lives are actually safe to use. It's a perfect blend of traditional craftsmanship and modern technology.

Technical Summary

Here is a detailed technical summary of the report "Automated Optical Inspection of Surgical Instruments" by Zunaira Shafqat, Atif Aftab Ahmed Jilani, and Qurrat Ul Ain.

1. Problem Statement

The surgical instrument manufacturing industry in Pakistan, particularly centered in Sialkot, is a major global exporter (contributing ~$359 million annually). However, the sector faces critical challenges:

• Quality Control Limitations: Current quality assurance relies heavily on manual visual inspection. This method is labor-intensive, prone to human error and fatigue, and inconsistent.
• High Stakes of Defects: Defects such as corrosion, pores, cracks, scratches, and rust can lead to surgical site infections (SSI), instrument failure during procedures, and severe patient harm.
• Economic Impact: Inconsistent quality leads to batch rejections, canceled export orders, and loss of reputation in international markets (EU/US).
• Data Scarcity: There is a lack of structured, annotated datasets specifically for surgical instrument defects to train robust AI models.

2. Methodology

The authors propose a comprehensive Automated Optical Inspection (AOI) system named SurgScan, utilizing deep learning and computer vision. The methodology follows a structured pipeline:

A. Data Curation and Preparation

• Collaboration: Data was gathered in partnership with industry leaders Dr. Frigz International and Daddy D Pro.
• Dataset Composition: The final dataset contains 4,414 images (2,738 defective, 1,676 non-defective) covering 11 distinct instrument types (e.g., Scalpels, Scissors, Forceps, Curettes, Probes).
• Defect Classes: The system targets four primary defect categories: Corrosion, Cuts, Scratches, and Pores.
• Acquisition: Images were captured using a Canon EOS 250D in a controlled environment (Puluz light box, green background) to ensure consistency.
• Augmentation: To improve generalization, techniques such as rotation (±20°), brightness/contrast adjustment (±20%), Gaussian noise addition, and unsharp masking were applied.
• Splitting: The data was stratified into an 80% training set and a 20% validation/test set.

B. System Architecture

The solution employs a two-stage classification approach:

1. Instrument Classification: A model first identifies the specific type of surgical instrument (e.g., distinguishing a scissor from a forceps). This is crucial because different instruments have unique geometries and surface textures that affect how defects appear.
2. Instrument-Specific Defect Detection: Once the instrument is identified, a specialized model trained specifically for that instrument type detects and classifies defects. This reduces false positives caused by structural features (e.g., natural pores or joints) being mistaken for defects.

C. Model Selection and Training

• Algorithms Evaluated: The team compared several state-of-the-art models: EfficientNet-b4, ResNet-152, ResNeXt-101, YOLOv8, and YOLOv11.
• Training Strategy: All models were trained on the same dataset with identical hyperparameters to ensure a fair comparison.
• Selected Model: YOLOv8 was selected as the optimal architecture due to its superior balance of real-time inference speed and high accuracy, which is critical for industrial throughput.

D. Implementation (SurgScan)

A web-based prototype was developed using a 3-Tier Client-Server Architecture:

• Frontend: React.js for the user interface.
• Backend: Django (Python) with Django Rest Framework (DRF) for business logic and API management.
• Database: PostgreSQL for storing user data, batch information, and inspection results.
• Image Processing: OpenCV and Pillow libraries for pre-processing (resizing, cropping, enhancing) before inference.

3. Key Contributions

• Industry-Academia Collaboration: Established a robust pipeline involving direct data collection and expert annotation from major Pakistani manufacturers (Dr. Frigz and Daddy D Pro).
• Specialized Dataset: Created a curated, annotated dataset of 4,414 images specifically for surgical instrument defects, addressing a gap in medical AI research.
• Two-Stage Detection Framework: Proposed and validated a hierarchical approach (Instrument ID $\rightarrow$ Defect Detection) that significantly improves accuracy over single-stage general models by accounting for instrument-specific structural nuances.
• End-to-End Prototype: Developed "SurgScan," a functional web application that integrates the AI models into a practical workflow for batch inspection, user management, and statistical reporting.

4. Results

The comparative analysis demonstrated that YOLOv8 outperformed other models across all metrics:

• Performance Metrics: YOLOv8 achieved a Testing Accuracy of 99.58%, Precision of 99.83%, Recall of 99.61%, and an F1-Score of 99.72% for the Dressing Forceps classification task.
• Comparison:
  • YOLOv8: ~99.5% Testing Accuracy.
  • YOLOv11: ~98.9% Testing Accuracy.
  • EfficientNet-b4: ~98.8% Testing Accuracy.
  • ResNet-152: ~85.0% Testing Accuracy (significantly lower, likely due to overfitting or architectural mismatch for this specific task).
• Efficiency: YOLOv8 provided the necessary speed for real-time industrial application, whereas deeper models like ResNet-152 were computationally heavier and less efficient for high-throughput scenarios.

5. Significance and Impact

• Patient Safety: By automating the detection of subtle defects (micro-cracks, pores, rust), the system significantly reduces the risk of Surgical Site Infections (SSI) and instrument failure during surgery.
• Economic Benefits for Pakistan: The solution offers a scalable way for Pakistani manufacturers to meet stringent international standards (ISO, FDA-GMP), potentially reducing export rejections and securing the country's position in the global $30 billion surgical instrument market.
• Operational Efficiency: Replaces slow, subjective manual inspection with a fast, consistent, and objective automated process, allowing sterilization units to handle higher volumes of instruments.
• Scalability: The web-based architecture allows for easy deployment in hospitals and manufacturing plants, providing a blueprint for AI-driven quality control in the medical device sector.

In conclusion, this report presents a successful transition from manual to automated quality control in the surgical instrument industry, leveraging deep learning to enhance patient safety and industrial competitiveness.