CT-based lung cancer detection using spatially localized integral transforms with U-NET classification

Abstract

Lung cancer remains the leading cause of cancer-related mortality worldwide, emphasizing the need for early, accurate, and scalable detection methods. Low-dose computed tomography (LDCT) has improved early diagnosis, yet challenges like image noise, low contrast, and subtle nodule features often limit reliable interpretation in large-scale screening. This paper proposes a computationally efficient computer-aided detection (CADe) framework that integrates a rotation-invariant, spatially localized integral transform feature extraction with a U-Net-based classifier to enhance lung nodule detection and segmentation. The approach strengthens spatial feature representation while maintaining low computational and memory demands, enabling real-time use in resource-limited clinical settings. Implemented in MATLAB and evaluated on the Cancer Imaging Archive (TCIA) dataset, the system achieved 99.32% classification accuracy, 88.88% specificity, 84.21% precision, 87.3% intersection over union (IoU), and 92.9% dice similarity coefficient (DSC). These results show clear improvements over conventional methods, particularly in rotational robustness and efficiency—key requirements for scalable screening. Although precision and IoU could be further optimized, the framework demonstrates strong potential for clinical adoption. By providing accurate, fast, and robust nodule analysis, this work advances practical high-performance tools for early lung cancer detection, especially in resource-constrained environments, ultimately contributing to better patient survival rates.