PERFORMANCE COMPARISON OF DEEP CNN ARCHITECTURES FOR LUNG AREA SEGMENTATION IN CHEST IMAGING

Abstract

Lung area segmentation is a critical preprocessing step in computer-aided diagnosis systems for respiratory diseases such as lung cancer and pneumonia. Accurate segmentation enhances the detection and monitoring of pathological conditions but manual delineation is time-consuming and subject to variability. This research aims to identify the most effective convolutional neural network (CNN) architecture for automated lung segmentation by comparing three models: U-Net, DeepLab, and a proposed hybrid model combining U-Net with ResUNet_Light. The models were trained and evaluated using a publicly available chest CT dataset under identical experimental settings, including preprocessing steps, training parameters, and standard evaluation metrics: Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Precision, and Recall. Results show that the proposed U-Net + ResUNet_Light model achieves the best performance across all metrics (DSC: 0.6767, IoU: 0.5652, Precision: 0.8480, Recall: 0.7920), outperforming both U-Net and DeepLab. These improvements are attributed to the integration of residual blocks, which enhance feature propagation and gradient flow, enabling better generalization and segmentation accuracy, especially along complex lung boundaries. In contrast, while DeepLab performs well in capturing contextual information, its higher complexity may hinder real-time applicability. U-Net, though efficient, showed limitations in accurately segmenting irregular regions. The findings demonstrate the potential of the proposed model for clinical deployment, where both accuracy and efficiency are critical. This study contributes to the development of more robust deep learning-based segmentation methods and highlights the importance of architectural enhancements in CNN design for medical image analysis.