Dust Particle Recognition using microscopic images
Background
The amount of cases of coal worker's pneumoconiosis, also known as black lung disease in underground coal mines in the United States, has significantly increased in the last two decades [1]. Different factors might be playing an important role, such as changes in dust characteristics, mining operations, and job duties [2]. To understand the problem and develop control strategies, a comprehensive knowledge of the dust and its constituents is critically needed. Current monitoring technologies allow quantifying the total mass concentration in underground mines in near real-time. Still, quick monitoring applications need to provide more timely information for a broader range of dust metrics.
Disease Resurgence
After decades of declining incidence of lung disease, an uptick was noted in the mid 1990s.
Possible contributing factors:
Shifting of specific dust characteristics.
Mining practices.
Very little data is available on this front to link mine exposure conditions with health outcomes.
The composition of dust is highly complex.
More information about specific dust constituents and enhanced monitoring is critically needed!
To fill the gap between the lack of information on dust characteristics and the need for more reliable-efficient methods to characterize dust in underground coal mines, we propose using machine learning techniques on optical microscope images of dust to allow quick identification and classification of respirable coal mine particles.
NOVEL SOLUTION
As a novel approach to this problem, we are investigating the efficacy of portable optical light microscopy and image processing techniques that incorporate machine learning algorithms as the basis of a new semi-continuous dust monitoring concept.
Previous Work
There has been a lot of work related to mineral classification using optical microscopy techniques and digital imaging. Optical properties have been widely used over time to characterize and classify minerals. Progress in imaging processing and machine learning techniques during the last few years allows the use of optical characteristics to achieve automatic mineral classification [4].
Here are some of the algorithms used in the literature for similar applications [5 - 10]
Decision Trees
Linear Discriminant
Feed-forward Neural Network
Nearest Neighbors
K-nearest Neighbors
Nearest Mode
Optimal Spherical Algorithms
Unsupervised Learning
Cascade Approach with Artificial Neural Networks
Ensemble Classifiers
Convolutional Neural Networks
Support Vector Machines
Random Forest
Logistic Regression
Multilayer Perceptron
Gaussion naive Bayes
Findings from prior research define three main dust sources of dust in underground coal mines:
Coal strata cutting
Rock dust applicaton
Rock strata cutting
Each source of dust has different particle types associated with it:
Coal
Silica
Kaolinite
Calcite or rock dust
In this work we will focus on classifying dust particles in one of these four categories. In summary, we are dealing with a four-class classification problem.
Data Collection
Sample Preparation
The dust was sampled using a 10-mm nylon cyclone attached to a pump working at a flow rate of 2.0 L/min.
The respirable fraction of the dust was deposited on a glass coverslip attached to the filter.
The glass coverslip with dust particles is taken and prepared for microscope imaging.
Microscope Imaging
All the images were captured under 4 different lighting conditions using an Olympus BX53M Polarizing Microscope.
Dataset
4 lighting conditions/frame
150 frames/mineral
A total of 600 frames
Total number of images: 2400
Each frame was stacked in 12 levels
4 lighting conditions x 3 color channels
Manual Feature Extraction Approach
The following features were manually extracted from the images:
Area
Eccentricity
Equivalent diameter
Extent
Length of major axis
Length of minor axis
Orientation
Perimeter
Solidity
Transmitted plane-polarized - Max particle intensity
Transmitted cross-polarized - Max particle intensity
Reflected plane-polarized - Max particle intensity
Reflected cross-polarized - Max particle intensity
Transmitted plane-polarized - Min particle intensity
Transmitted cross-polarized - Min particle intensity
Reflected plane-polarized - Min particle intensity
Reflected cross-polarized - Min particle intensity
Transmitted plane-polarized - Mean particle intensity
Transmitted cross-polarized - Mean particle intensity
Reflected plane-polarized - Mean particle intensity
Reflected cross-polarized - Mean particle intensity
Results
After training different models on the dataset, the accuracy was calculated to measure the performance
The best model was the radial basis function support vector machine algorithm with 90.80% accuracy. The confusion matrix for that model is shown on the right. The model does a good job separating coal versus other minerals. That makes sense, since the optical properties of the coal are expected to be very different. When comparing the performance between mineral particles, we see more misclassification. The mineral particles are more difficult to distinguish, since the optical properties are more similar.
Convolutional Neural Network (CNN)
CNNs have become very popular for working with images since 2012, after which CNNs have been involved for every winning model for the ImageNet competition held annually.
We make use of a fairly simplistic CNN model with 2 convolutional layers, each followed by a max pooling layer. the output of the convolutional layer is passed onto a layer of neurons to flatten the image dataset, and then finally a dense layer of size 4 where each neuron represents a class.
The reason for using a simple CNN model is two-fold:
Given the nature of the problem, we do not expect the model to work with very complex features, and keeping the model simple avoids overfitting even with the relatively small number of training samples we had collected.
A simple model requires less computational resources and thus converges faster.
(The tensor flow package in python to generate and train the CNN models)
Initial models
Even with the simple model, we had overfitting issues with our initial dataset. Our initial dataset consisted of 48 images per class, which meant that in reality we had only 12 images per class since the images were duplicated 4 times under different lighting conditions.
To counter this issue, we collected a larger dataset and explored image augmentation techniques that would further increase our training data size
Image Augmentations
Cropping
We use an overlapping windows to crop our original image repeatedly. This not only increases the number of samples in training, but also reduces the number of features thus makes a large impact on the error due to bias.
Data size multiplier: 8
Vertical flips
We flip the original image along the vertical axis to double our training data size.
Data size multiplier: 2
Rotations
We can rotate our image by 90 degrees to get new images that would definitely have the same classification.
Data size multiplier: 4
CNN results
It comes as no surprise that the highest training data size resulted in the highest accuracy. We were unable to run tests with rotations as the data size increased beyond the capacity of the machines we were using. Here's the confusion matrix for the best model.
The confusion matrix below describes the result from the best model
Conclusion and future work
We have shown that we can definitely classify dust particles using ML methods and microscopic images. Having said that, the current models are trained on lab-generated pure samples and will have a much higher error if the sample collection mechanism is not as strict. If this methodology is to be used under real life conditions, it is recommended that models be trained on samples collected from a variety of locations.
The CNN models proved to be generally more predictive where the best CNN model had a 99% accuracy. The models running on manual feature selection, however may prove useful due to their interpretability and would likely help further research. CNN models also used only one lighting condition and thus are the best candidate for systems where particle detection is the only goal.
Manual feature extraction showed lower performance, Nevertheless, this model is very easy to interpret and can serve as a guidance to understand which features might be playing important roles in terms of prediction power. In this analysis, only linear features were explored. More work can be done using a wider variety of non-linear predictors that eventually could help improving the overall performance of the model.
The feature extraction can be automated using a more robust approach. For instance in this paper [10] they use the Inception-v3 architecture to extract the features from the images and then use those as inputs for different machine learning techniques to evaluate the classification accuracy. This may lead to a less biased model, since we are not prone to leave some predictive power in the table by manually selecting the features.
The limited number of images and the time that it requires to obtain a dataset that captures the variability can be challenging. Synthetic data generation can be a good tool to overcome this limitation. Based on the current images, more images can be synthetically generated with different variance to train the models. More work in this direction needs to be done, but it is definitely something that can bring benefit exploring.