The challenges with detecting pneuoconiosis
Pneumoconiosis is caused by long-term inhalation of respirable dust, such as coal, asbestos and silica. It is more commonly known as black lung. About 25,000 people died of pnemoconiosis globally in 2013. In Queensland, Australia, about 145 cases of mine dust lung diseases have been diagnosed since 1984. But with the recent re-emergence of pnemoconiosis, more cases are feared to have been missed. Poor dust control and patchy medical screening are to blame for the resurgence of this potentially deadly disease even in developed countries.
To date, there has been a lack of systematic, automated and objective systems for detecting the presence and assessing the progression of pnemoconiosis for individual coal miners other than by expert radiologists.
Deep learning has become very popular and has been used practically in many industry domains. However, one common barrier for deep learning to solve real-world problems remains the amount of labelled training data. In practice, imbalanced datasets often come up with majority of training data from a single class and limited number of training samples from another class. This can lead to biased prediction in favour of the majority class.
A cascade learning architecture for pnemoconiosis detection
For pneumoconiosis detection, we have abundant training data for normal x-rays, however, the number of X-rays with signs of pneumoconiosis is limited. To address this issue, we propose a cascade learning architecture for the automated pneumoconiosis detection.
Machine learning based lung field segmentation - we used a pixel-based machine learning algorithm that employs Pixel Classification (PC) to distinguish between lung and non-lung areas in a radiograph.
Cycle-Consistent Adversarial Networks (CycleGAN) image generator - we trained a CycleGAN using 56 normal and 56 pneumoconiosis lung fields to generate 1000 normal and pneumoconiosis lung field images, respectively.
Image augmentation - all images are normalised. For training images, their mean is set to zero, they are also divided by their standard deviation, randomly zoomed, flipped horizontally. Their pixels intensities are also sheared.
Convolutional Neural Network (CNN) based image classifier - the classifier is composed of 18 neural network layers, including eight convolutional layers to extract feature maps, four pooling layers, three dense layers, two dropout layers, and the last layer using sigmoid activation function to output probability scores for each class - normal and pneumoconiosis.
Prevention is the key to management
Pneumoconiosis is incurable, and prevention is the key to management. Early detection of pneumoconiosis through routine medical screening is critical to preventing complications including death. We hope that this technology can be used for the pre-screening of occupational lung diseases, and to address the issues of variability in identifying pneumoconiosis, and the shortage of B-readers.
The cascade learning model can be potentially used in other medical imaging applications when a training dataset is imbalanced or lack of diversity. The proposed method outperforms others and achieves a sensitivity of 93.33 per cent, a specificity of 88.46 per cent and an overall accuracy of 90.24 per cent.
After developing a deep learning based automated diagnostic tool for pneumoconiosis detection from chest radiographs, with a very promising sensitivity of 93 per cent and accuracy of 90 per cent, we are looking at making our tool more robust and suitable for pre-screening of pneumoconiosis.
Following being granted additional funding from the Coal Services Health and Safety Trust , we will now collect, nationally and internationally, a set of chest radiographs with different stages of disease and expand our current tool to become a multi-class grading system. Additionally, we will set up a pilot study where our tool will be trialled in a clinical setting alongside human readers.
Our highly skilled team of world class researchers and engineers is open to partnerships and collaborations for research, development and commercialisation.