<![CDATA[Background: Lung cancer is the leading cause of cancer-related mortality globally, with over 2.2 million new cases and 1.8 million deaths reported annually (WHO, 2022). Pulmonary nodule detection through low-dose computed tomography (LDCT) screening is the most effective method for early lung cancer identification. However, automated systems still face significant challenges: high false positive rates, limited sensitivity for micronodules (<5 mm), and poor segmentation accuracy for nodules with irregular morphology or juxtapleural attachment. Objective: Lung nodules early discovery is key to treating lung carcinoma, but even conventional systems' micronodules still have high false positives and low accuracy. Method: This study presents an end-to-end hybrid pipeline that uses the LUNA16 database to tackle this issue. The initial stage is to make use of YOLOv12 for Region of Interest (ROI) extraction, with 3D-CNN carrying out false positive filtering through volumetric verification as a gate. The final phase conducts pixel-level precision segmentation using Adaptive Bayesian Fusion on U-Net Residual 3D ensemble (local texture features) and ViT-UNETR (global anatomical context). Results: Experiments showed superior performance level 99.99% Accuracy, Mean Dice Similarity Coefficient (DSC) at 93.88% and IoU is 90.45%. The system was very robust, reaching 97.33% DSC in the micro nodule category (<5 mm). Conclusion: In summary, this integrated architecture delivers an objective, efficient and high-quality solution for automated Diagnosis.]]>
