Pneumonia Detection From X-Ray Images With Deep Convolutional Neural Network

• Main Author: Nur Syaza Nabilah Binti Mohd Shukri
• Language: English
• Published: 2023
• Subjects: CNN; Deep Learning; Covolutional Neural Network

Pneumonia is an inflammatory lung disease that can be considered as a fatal disease. Normally, this lung infection can only be diagnosed by radiologists. The disease can be detected through conducting X-ray scans of patients’ chests by Magnetic Resonance Imaging (MRI) means or CATSCAN. The problem that we are currently facing today is that radiologist use chest X-rays to detect pneumonia is not an easy task. The pneumonia appearance in X-ray images are usually cannot be seen very clearly. To mitigate this, deep learning method is introduced. Deep learning method has been used for many things such as identifying objects in images, image classification, handwritten recognition and many more. Deep learning methods also have recently become popular for analyzing medical images. In this project, deep learning is used to detect pneumonia by developing a convolutional neural network (CNN) model which is trained to recognize the disease using frontal chest X-rays images as valuable data. This project used a dataset obtained from Kaggle. This dataset consists of 3 folders which are train set, test set and validation set. Each folder is divided into two subfolders namely as pneumonia and normal. These chest X-ray images were obtained from Guangzhou Women and Children’s Medical Center, Guangzhou. In this project, the CNN model trained with different cases of hyperparameter on the same dataset to investigate the best network architecture and hyperparameter combinations for the best classification accuracy. The hyperparameter cases are number of convolutional layers, number of filter and kernel size. The final results obtained are training accuracy of 0.9676, validation accuracy of 0.9324 and testing accuracy of 68.75% .This highest accuracy is obtained when using 5 convolutional layers, 3x3 kernel size and 64 as the number of filter. The highest testing accuracy obtained is 75%.