Akademik Veri Yönetim Sistemi
9th International Artificial Intelligence and Data Processing Symposium, IDAP 2025, Malatya, Türkiye, 6 - 07 Eylül 2025, (Tam Metin Bildiri)
Lung cancer is the leading cause of cancer-related deaths worldwide, making early diagnosis from CT images critically important for patient outcomes. In this study, a unique dataset was collected at Inönü University Turgut Özal Medical Center, consisting of 3526 CT slices from 210 patients manually annotated by expert physicians. The lesions, classified into benign, malignant, and cystic categories, were segmented using seven different deep learning architectures: U-Net, VGG16UNet, EFF-UNet, UNet++, FPNet, PSPNet, and EFF-PSPNet. The models were evaluated based on metrics such as Dice, Jaccard, precision, recall, and FPS. The most successful results were achieved by the EFF-UNet (Dice: 92.14%) and U-Net (Dice: 92.25%) models. While cystic lesions were detected with high accuracy, benign lesions presented the lowest results due to their morphological variability. Visual analyses indicated that some models exhibited uncertainty in class differentiation, and the quality of annotations was found to directly influence model performance. This study demonstrates the potential of deep learning-based segmentation for clinical decision support systems and contributes to the field with its original dataset.