DEVELOPMENT OF CONVOLUTIONAL NEURAL NETWORK-BASED COMPUTER AIDED DIAGNOSIS FOR EARLY DETECTION OF LUNG ABNORMALITIES WITH ULTRASOUND IMAGES

• Main Author: Immanuel, Matthew
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85680
• Institution: Institut Teknologi Bandung

The lungs are an organ that has a very important role in the respiratory system, and lung abnormalities can have a negative impact on the health of the body. Pneumonia is one of the major health problems worldwide, including in Indonesia, especially in children. Computed Tomography-scan (CT-Scan) is often used in hospitals to detect abnormalities in the lungs. However, CT-Scan has limited availability, and its radiation exposure can damage body systems, especially in children and pregnant women. Ultrasound (US) is a safer alternative medical imaging method as it does not use ionizing radiation. Other advantages of ultrasound are its wide availability, more affordable examination costs, and its portable nature. This study aims to develop a Convolutional Neural Network (CNN)-based Computer Aided Diagnosis (CAD) for the classification of lung abnormalities based on recorded ultrasound images or lung ultrasound (LUS). The CNN model is developed through a series of stages that include data pre-processing, CNN architecture selection, and CNN model optimization settings. The data pre- processing used includes image resizing, data normalization, and data augmentation. CNN architecture selection includes the use of basic CNN, VGG-16, ResNet-19, and GoogleNet-V3 as pre-trained CNN. CNN model optimization is done through setting the optimization algorithm, learning rate, mini-batch size, and epoch. The test and evaluation of these features were carried out using LUS images. The LUS CNN model using the best features has an accuracy of 92.33% and F1-Score of 0.97, with an indication of minimal overfitting and underfitting. These results indicate that the model is feasible to use for classifying LUS images. In addition, a CAD application has been developed to facilitate the use of the CNN LUS model in detecting lung abnormalities based on LUS images. Keywords: lung abnormalities, ultrasonography, CAD, pre-processing, CNN