SEARCH RESULT

This study introduces a novel convolutional neural network (CNN) architecture integrating multi-scale dilated convolutions with a sequential dense network for improved lung disease detection in computed tomography (CT) scans. Dilated convolutions increase the receptive field without adding extra parameters, enabling the model to capture spatial features from various scales. The proposed model effectively detects subtle patterns associated with lung conditions such as cancer, benign tumors, and normal tissues. Evaluated on the IQ-OTH/NCCD dataset containing 1,190 CT images across three classes, the model achieved its highest accuracy of 95.33% with a dilation rate of 32, also showing high sensitivity (95.33%) and specificity (97.67%). Comparative analysis across multiple dilation rates (4, 8, 16, 32, 64, 128) showed that mid-range dilation values yield optimal performance, while excessive dilation leads to diminished returns. To mitigate class imbalance, geometric data augmentation techniques were employed. Unlike traditional CNNs, our approach balances computational efficiency and high classification performance, making it suitable for real-world deployment in resource-constrained settings. The model’s ability to accurately classify small and dispersed abnormalities enhances its potential in early diagnosis. In summary, the proposed multiscale dilated CNN offers a robust and efficient solution for CTbased lung disease detection, demonstrating promising results for integration into clinical decision support systems.

International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES

Aleyna YERLİKAYA Alper Talha KARADENİZ Zafer Cömert