An accurate black lung detection using transfer learning based on deep neural networks

Abstract

Coal Worker Pneumoconiosis (CWP), commonly called Black Lung (BL), is an incurable respiratory disease caused by long-term inhalation of respirable dust. Privacy restrictions and disease incidence placed limits on the available BL datasets, which introduces significant challenges for training deep learning (DL) models. Recently, transfer learning has been seen as an efficient DL method for automatic disease detection with small datasets. This paper investigates BL detection in chest X-rays using transfer DL knowledge from a CheXNet model on a small dataset. A training image set of real, segmented lung X-ray images, with and without BL, was used as a benchmark for detection accuracy. The training data set was then augmented using a Cycle-Consistent Adversarial Networks (CycleGAN) and Keras Image Data Generator, to generate training data with real, augmented and synthetic CWP radiographs to the CheXNet model (with and without pre-trained weights). The effects of different dropout nodes as a blocking factor was also investigated. The accuracy, sensitivity (recall or true positive rate), specificity (true negative rate) and error rate (ERR or incorrect prediction rate) using 3-fold cross-validation experiments was compared for each transfer learning experiment. The total execution time for binary classification of our model also measured. While no definitive conclusion could be reached regarding the effect of dropout rates, results indicated an improvement of classification accuracy from transfer learning.