SegNet-Based Framework for Robust Lung Tumor Segmentation from PET/CT Scans

Abstract

Lung cancer remains one of the leading causes of cancer-related mortality worldwide, often due to latestage diagnosis and the complexity of tumor localization in thoracic imaging. Accurate and automated segmentation of lung tumors from PET/CT images is essential for early diagnosis, treatment planning, and outcome monitoring. Manual segmentation is time-consuming and prone to observer variability, underscoring the need for reliable deep learning-based solutions. This study proposes an automated lung tumor segmentation framework using the SegNet architecture, a deep encoder-decoder convolutional neural network. A dataset of 1, 425PET/CT images, manually annotated by expert radiologists, was utilized. Data augmentation techniques were applied to improve generalization. SegNet was trained to perform pixel-wise binary classification, and its performance was benchmarked against the widely used U-Net model. Evaluation metrics included Accuracy, Recall, Dice coefficient, and Intersection over Union (IoU).The proposed SegNet model achieved strong segmentation performance across independent experiments. Average results were: Accuracy of 92.24% ± 1.42, Recall of 94.02% ± 1.287, Dice coefficient of 93.47% ± 1.4, and IoU of 93.03% ± 1.2. Compared to U-Net (Dice: 92.18% ± 1.081, IoU: 91.70% ± 1.287 ), SegNet demonstrated improved spatial boundary accuracy, particularly for tumors located near complex anatomical structures. Statistical tests confirmed the significance of the performance difference ( p < 0.05 ). The SegNet-based model provides accurate and robust segmentation of lung tumors in PET/CT images, outperforming U-Net under the same conditions. Its use of max-pooling indices enhances spatial precision, making it well-suited for clinical applications. Future work will explore 3D extensions, multi-class segmentation, and multi-center validation to enhance its applicability in real-world diagnostic workflows.