Lung Nodule Detection and False Positive Reduction Using Multi-View and Background Images

This paper presents a lung nodule detection and false positive reduction approach based on YOLOv8x, multi-view CT representations, and background slices used as negative samples. Lung cancer remains one of the leading causes of cancer-related mortality worldwide, and early detection is essential to improve patient survival. Although computed tomography (CT) provides detailed three-dimensional information about the lungs, the large number of generated slices increases the complexity of radiological analysis and can contribute to diagnostic errors. To address this challenge, Computer-Aided Detection (CAD) systems have been developed to assist radiologists in identifying suspicious regions. In this work, we combine axial, coronal, and sagittal representations in a single YOLOv8x detection pipeline, use additional LIDC-IDRI annotation clusters with fewer than three radiologist markings as auxiliary positive samples, and introduce nodule-free background slices as negative training samples. Detections are consolidated through intra-view clustering and weighted inter-view fusion. Experiments following the LUNA16 cross-validation protocol show that the proposed strategy improves detection performance while substantially reducing false positives. The axial baseline achieved a Competition Performance Metric (CPM) score of 0.553 with 32.61 candidates per scan, whereas the multi-view configuration with a nominal budget of 20,000 background images achieved a CPM score of 0.863 and reduced the average number of candidates to 9.11 per scan while preserving high sensitivity. These results show that the combined use of multiple anatomical views and background-negative training can substantially reduce spurious candidates in CT-based lung nodule detection.