M.S. AAI Capstone Chronicles 2024

CNN Lung Disease Classification

8

CNNs are exceptionally well-suited to this medical imaging project due to their inherent strengths in feature extraction, spatial awareness, and efficient handling of visual data. As demonstrated in work by Krizhevsky et al. (2012), CNNs excel at automatically learning and extracting intricate features from images, crucial for identifying subtle patterns and anomalies in medical images like X-rays. Their hierarchical structure enables the capture of both low-level features (e.g., edges, textures) and high-level features (e.g., organ shapes, disease-specific patterns), making them highly effective for detecting and classifying diseases within the NIH Chest X-ray Dataset (White et al., 2023). Furthermore, CNNs preserve the spatial relationships between pixels, a critical aspect of medical image analysis where the location and arrangement of features are essential for accurate diagnosis. For instance, in chest X-rays, the spatial context of the heart, lungs, and blood vessels is crucial for identifying abnormalities (Krizhevsky et al., 2012). CNNs retain this context throughout the analysis, leading to more accurate and reliable results. Several studies have demonstrated the effectiveness of building Convolutional Neural Networks (CNNs) from scratch using large medical datasets. A notable study by Dahmane et. al. (2019) involved a hybrid approach combining a CNN built from scratch with a pre-trained VGG19 model to detect pneumonia in chest X-ray images. Their custom-designed CNN architecture, when combined with the VGG19 model for feature extraction, achieved over 99% accuracy on the Guangzhou Women and Children's Medical Center dataset. This hybrid approach, which leverages both the strengths of pre-trained models and custom-designed CNNs, is something we explored in our research to further optimize our model's performance​.

166