Accurate pulmonary nodule detection via progressive multi-scale networks and geometric constraint refinement

Pulmonary nodule detection in CT images plays a crucial role in the early diagnosis of lung cancer. However, the development of efficient detection systems, especially in reliable detection of small nodules with low false positive rates, remains challenging as small nodules often occupy minimal pixel regions and are morphologically confusable with vascular and surrounding anatomical structures. In this paper, we propose an alternate nodule detection framework to solve such tricky issue, which exhibits several salient features. Firstly, we propose a Progressive Multi-Scale Network (PMSN), a feature extraction module with a four-branch interactive architecture, specifically designed to enhance small nodule detection. By integrating this module into the YOLO object detection framework, we develop a PMSN-YOLOX network that effectively generates small-nodule-sensitive 2D candidates. Based on this, we develop a Geometric Constraint False Positive Reduction (Geo-FPR) algorithm via spatial clustering and confidence scoring to eliminate inconsistent detections while preserving true nodules. Additionally, we employ a self-supervised pre-training framework and a dynamic segmentation enhancement strategy to further improve model robustness. Experimental results on the LUNA16 dataset demonstrate that the proposed method exhibits superior performance in small nodule detection and false positive reduction, achieving 100% sensitivity with 2 to 8 false positives per scan. Even at the strict 1/8 false positives per scan threshold, sensitivity remains 98.3%. In addition, only 1.2 s is required to process each case. In summary, the proposed approach significantly outperforms existing SOTA techniques in terms of both performance and efficiency.