Cystic fibrosis (CF) is a genetic disease that attacks multiple organs. It is an incurable illness in the respiratory system, with about one-sixth of native Australians suffering from it. Cystic Fibrosis Community Care* shows that 1 in 25 people in Australia are carrying defective CF genes and nearly 90 babies each year are born with CF.
Clinicians diagnosed the pulmonary CF mainly through chest radiographs (CXRs). This study proposes a novel computer-aided diagnostic (CAD) framework with deep learning features for automated scoring of CF in CXRs. Within this framework, key components and algorithms are developed, examined and refined to achieve the best scoring performance as compared to human observers.
We present a framework to analyse chest radiographs for cystic fibrosis using machine learning methods. We compare the representational power of deep learning features with traditional texture features. Specifically, we respectively employed VGG-16-based deep learning features and Tamura- and Gabor-filter-based textural features to represent the cystic fibrosis images. We demonstrate that VGG-16 features perform best, with a maximum agreement of 82%. In addition, Tamura features’ accuracy of Tamura reaches 78%.
We also present a framework to analyse chest radiographs for cystic fibrosis using deep learning methods. We compare the effectiveness of two typical deep convolution neural networks, ResNet-152 and VGG-16, both of which are transferred the knowledge learned from the ChestX-ray14 dataset. We had the best scoring accuracies of 91% for ResNet-152 and 86% for VGG-16, 12.3% and 3.6% respectively better than the networks pre-trained by natural images on ImageNet. The overall best performance was exhibited by transfer-learning-based ResNet-152 with random forest classifier. It achieved 91.5% accuracy for classification across three levels of scoring (10, 15 and 20).