EMPLOYING DEEP LEARNING TECHNIQUES FOR CLASSIFICATION AND DETECTION OF PEDIATRIC LUNG DISEASE USING X-RAY IMAGES

Abstract

World Heal Organization (WHO) statistics place pneumonia as the prevalent cause of death for children under five years old, overtaking HIV, malaria, and other diseases. According to the WHO, one half to two-thirds of the global population lacks access to radiological services. Deep learning techniques can automatize lung disease detection. In this thesis, deep learning model architectures namely convolutional neural network (CNN) and vision transformers (VT) are used to construct models via transfer learning methodology. Models are trained and tested on two different datasets: the open-source Pediatric dataset and the Kazakhstan dataset. The best-obtained accuracy for the open-source dataset is 96.8% with a 97.7% f1 score for pneumonia using the vision transformer architecture. The fusion model consisting of both CNN and VT was constructed and found to be the best to classify the unseen another dataset source with 86.2% accuracy and 88.7% f1 score for pneumonia. The YOLO_V5 model applicability in multiclass detection of child lung diseases on chest X-ray images was tested using the Vinchext dataset and the Kazakhstan dataset. YOLO_V5 model was able to detect with moderate accuracy only the prevalent disease class boundaries in chest X-ray images.