Detection of Foreign Matter in Transfusion Solution Based on Gaussian Background Modeling and an Optimized BP Neural Network

Fuqiang Zhou, Zhen Su, Xinghua Chai, Lipeng ChenSensors 2014 · DOI

TL;DR

This paper introduces a novel automated system for detecting foreign matter in transfusion solutions within plastic bottles. By combining a specialized spin-stop mechanism, mixed illumination, Gaussian background modeling, and an optimized BP neural network, the system achieves high accuracy and speed in distinguishing foreign matter from bubbles, crucial for ensuring product safety.

What Problem Does This Paper Address?

Ensuring the purity of medical transfusion solutions is critical for patient safety, as foreign matter can lead to severe health consequences. Traditional manual inspection methods in China are subjective, prone to errors, and lack repeatability. While automated visual inspection systems exist, they often struggle with the complex visual characteristics of plastic bottles (e.g., surface reflections, embossed labels) and differentiating foreign matter from naturally occurring bubbles. This paper tackles the challenge of reliably detecting and classifying foreign matter in these complex environments, aiming for an objective, accurate, and efficient automated solution.

Methods

The proposed method employs a multi-stage approach:

  • Image Acquisition: A "spin-stop" mechanism is used to create a vortex in the transfusion solution, causing foreign matter to move. Mixed illumination (bar-shaped LEDs and a round condensing LED) is applied to enhance contrast between moving objects and the static background, making foreign matter appear as dark shapes.
  • Moving Object Detection: Gaussian background modeling is utilized to segment moving objects from the static background. This technique is robust to lighting changes and complex backgrounds, outperforming traditional frame differencing methods.
  • Feature Extraction: For each detected moving object, a set of fifteen features are extracted. These include geometric invariants (like Hu moments, area, length, width) and gray-level statistical properties (mean, variance, entropy).
  • Feature Selection: The ReliefF algorithm is employed to assign weights to these features, identifying those most discriminative for classification. This step aims to reduce dimensionality and improve the efficiency of the subsequent classification.
  • Classification and Optimization: A Back Propagation (BP) neural network is used to classify detected objects as either foreign matter or bubbles. To overcome the BP network's limitations of slow convergence and tendency to fall into local minima, a Mind Evolutionary Algorithm (MEA) is used to optimize its connection weights and thresholds, leading to improved accuracy and faster convergence.

Datasets

The paper describes experiments using samples of 250-mL plastic transfusion bottles. For training and testing, they used:

  • 300 bottles with typical foreign matter.
  • 100 bottles without foreign matter (used to obtain bubble samples).
  • A total of 200 bottles (150 with foreign matter, 50 without) were used for the overall system performance test.

The paper does not introduce a new benchmark dataset but rather uses a collection of real-world samples.

Key Insights and Ideas

  • Illumination is Key: The mixed illumination strategy is crucial for overcoming the challenges posed by reflective plastic bottles, creating high-contrast images of foreign matter.
  • Gaussian Modeling for Robustness: Gaussian background modeling proves more effective than simple frame differencing for detecting moving objects in complex and noisy environments, such as those found in transfusion bottle inspection.
  • Feature Engineering Matters: A combination of geometric and gray-level features, particularly Hu moments, are highly effective in distinguishing between different types of foreign matter and bubbles.
  • Optimization is Essential: Optimizing the BP neural network with MEA significantly improves both classification accuracy and convergence speed, making the system more practical for real-time applications.
  • Bubbles are the Main Challenge: The paper highlights that distinguishing bubbles from foreign matter is a primary challenge, and the proposed method, especially with MEA optimization, shows promise in this regard.

Key Results

  • The proposed method can effectively detect visible foreign matter in 250-mL transfusion bottles.
  • The misdetection rate and false alarm rate are low.
  • The overall system misdetection rate is 1.096%, and the false alarm rate is 1.250%.
  • The unqualified samples (foreign matter) are detected with accuracies of 97.833% (rubbers), 98.714% (hairs), and 97.640% (floating debris).
  • Qualified product detection accuracy is 98.75%.
  • The BP algorithm optimized by MEA achieved a testing accuracy of 98.76% with five hidden neurons, outperforming the standard BP algorithm (95.68% with ten hidden neurons).
  • MEA optimization also significantly reduced training time.

Limitations and Future Work

  • Misdetection Causes: The paper identifies two main reasons for misdetection and false alarms:
  • Changes in bubble brightness can be falsely identified as foreign matter.
  • Incomplete segmentation of moving objects can lead to errors in classification.
  • Future Work: The authors suggest future work will focus on:
  • Designing mechanisms to reduce bubble formation during the process.
  • Developing more effective features to better distinguish between bubbles and foreign matter.

Summary updated

Abstract

This paper proposes a new method to detect and identify foreign matter mixed in a plastic bottle filled with transfusion solution. A spin-stop mechanism and mixed illumination style are applied to obtain high contrast images between moving foreign matter and a static transfusion background. The Gaussian mixture model is used to model the complex background of the transfusion image and to extract moving objects. A set of features of moving objects are extracted and selected by the ReliefF algorithm, and optimal feature vectors are fed into the back propagation (BP) neural network to distinguish between foreign matter and bubbles. The mind evolutionary algorithm (MEA) is applied to optimize the connection weights and thresholds of the BP neural network to obtain a higher classification accuracy and faster convergence rate. Experimental results show that the proposed method can effectively detect visible foreign matter in 250-mL transfusion bottles. The misdetection rate and false alarm rate are low, and the detection accuracy and detection speed are satisfactory.
BibTeX
@inproceedings{zhou2014detection,
  author = {Fuqiang Zhou and Zhen Su and Xinghua Chai and Lipeng Chen},
  title = {{Detection of Foreign Matter in Transfusion Solution Based on Gaussian Background Modeling and an Optimized BP Neural Network}},
  booktitle = {Sensors},
  year = {2014},
  doi = {10.3390/s141119945},
}

Loading related content...

Figure