Abstract: Pulmonary nodules in low-dose computed tomography screening are often small, ill-defined, and easily confused with vessels or pleural attachments, leading to unstable segmentation, excessive false positives, and whole-patient risk prediction that is vulnerable to candidate noise. To address these issues, this study proposes a chained deep-learning framework. First, GAU-FU-Net is developed on the LIDC-IDRI dataset by introducing global-attention upsampling and three-path feature fusion into the U-Net decoder to improve the segmentation of small nodules and weak boundaries. Second, candidate purification is performed using high-confidence reference masks and a dual-threshold matching strategy, followed by a lightweight convolutional neural network(CNN) for false-positive reduction. Finally, the front-end modules are transferred to DSB2017, where 2.5D candidate bags are constructed and a multiple-instance learning model with ResNet instance encoding and Top-K aggregation is used for whole-patient malignancy risk prediction. Experimental results show that GAU-FU-Net achieves a Dice score of 0.8027 on LIDC-IDRI, representing an absolute improvement of 1.80 percentage points over U-Net. On the independent DSB2017 test set, the whole-patient risk prediction model attains an AUC of 0.9683 and an AP of 0.9040. These findings indicate that candidate purification and false-positive suppression can effectively improve the reliability of whole-patient risk quantification under weak supervision.