Fig 1.
The proposed MLND-IU framework for lung nodule detection (generating candidate regions by improving RetinaNet high recall in the first stage, refining features by combining the channel-space attention mechanism of AG-UNet++ in the second stage, and dynamically suppressing false positives by fusing multi-slice information using 3D context pyramid in the third stage to realize multi-stage accurate detection of lung nodules).
Fig 2.
Enhanced Feature Pyramid Network based on cross-scale feature interaction.
Fig 3.
Attention-directed UNet++-based feature refinement.
Fig 4.
The designed 3D-CPM Module.
Table 1.
Introduction to experimental platforms and datasets.
Table 2.
Configuration summary covering backbones, input sizes, optimizer, and memory footprint.
Fig 5.
Comparison of feature generation results for candidate regions of interest (ROIs) before and after data augmentation preprocessing.
The preprocessed image (right panel) demonstrates significantly enhanced edge details (e.g., clearer delineation of spiky and lobular structures). Contrast enhancement accentuates the texture of 3–5 mm sub-centimeter nodules, overcoming the boundary blurring issue in the original image (left panel) caused by low signal-to-noise ratio.
Fig 6.
Training process of the MLND-IU model.
Table 3.
Comparison of imbalance-aware losses on sub-centimeter nodule detection.
Table 4.
Ablation results of the proposed MLND-IU model.
Table 5.
Ablation isolating the cross-scale module (without DFL).
Fig 7.
Sensitivity analysis of key hyperparameters β.
Fig 8.
Ablation study with 95% confidence intervals and statistical significance.
Table 6.
Quantitative comparison of DABM with common attention modules.
Fig 9.
Computational Profile of the MLND-IU Pipeline.
Fig 10.
Comparison of detection results for <3mm nodules across state-of-the-art models.
Green contours indicate true positives (TP), red contours indicate false positives (FP), and yellow arrows highlight regions of interest (e.g., vascular adhesion artifacts).
Fig 11.
Comprehensive performance comparison: (A) Sensitivity vs. FP/Scan trade-off; (B) Radar chart of multi-dimensional metrics; (C) Pareto efficiency frontier; (D) Distribution of Dice coefficients across models.
Fig 12.
Size-specific sensitivity versus FROC curve analysis.
Fig 13.
Comparison of different attention modules integrated into our framework.
Table 7.
SOTA performance evaluation on LIDC-IDRI dataset (+3.9% (Recall), + 25.1% (Precision), −65.9% (FP/Scan)).
Fig 14.
Comprehensive comparative analysis of multi-dimensional model performance.
(A) Radar chart comparing the performance balance of MLND-IU versus mainstream baseline models across six key dimensions, all metrics normalized (higher values indicate better performance), (B) The scatter plot illustrates the distribution of more advanced models within the “accuracy-efficiency” trade-off space. Error bars indicate the interquartile range (IQR) of the Dice coefficient. The red dashed line marks the clinical real-time processing threshold (3 seconds per case), while the green region represents the ideal “high-accuracy-low-latency” quadrant. MLND-IU (starred points) demonstrates the best overall performance in both analyses.
Fig 15.
Performance generalizability for cross-data detection.
Fig 16.
Inference speed on different hardware platforms.