Abstract
Accurate detection of pulmonary nodules with high sensitivity and specificity is essential for automatic lung cancer diagnosis from CT scans. Although many deep learning-based algorithms make great progress for improving the accuracy of nodule detection, the high false positive rate is still a challenging problem which limited the automatic diagnosis in routine clinical practice. In this paper, we propose a novel pulmonary nodule detection framework based on a 3D Feature Pyramid Network (3DFPN) to improve the sensitivity of nodule detection by employing multi-scale features to increase the resolution of nodules, as well as a parallel top-down path to transit the high-level semantic features to complement low-level general features. Furthermore, a High Sensitivity and Specificity (HS$^2$) network is introduced to eliminate the falsely detected nodule candidates by tracking the appearance changes in continuous CT slices of each nodule candidate. The proposed framework is evaluated on the public Lung Nodule Analysis (LUNA16) challenge dataset. Our method is able to accurately detect lung nodules at high sensitivity and specificity and achieves $90.4\%$ sensitivity with 1/8 false positive per scan which outperforms the state-of-the-art results $15.6\%$.
Introduction
Computer-Aided Diagnosis (CAD) systems are increasingly vital in assisting radiologists with the early detection of lung cancer. However, a significant bottleneck persists: high false positive (FP) rates. To ensure clinical safety, models must maintain high sensitivity, but this often comes at the cost of numerous false alarms—sometimes up to 8 per scan. This leads to alert fatigue and limits clinical adoption.
The authors identify two primary challenges:
- Morphological Similarity: Normal anatomical structures, such as blood vessels or lesions, can appear nearly identical to nodules in individual 2D CT slices.
- Scale Discrepancy: Pulmonary nodules vary significantly in size (3mm to 30mm) and occupy a minuscule fraction of the total CT volume, making consistent detection difficult for standard architectures.
Methodology
The proposed 3DFPN-HS² framework utilizes a two-stage pipeline to decouple candidate generation from false positive reduction.
3D Feature Pyramid Network (3DFPN)
To address scale discrepancy, the authors adapted the 2D Feature Pyramid Network for 3D volumetric data. The 3DFPN processes 3D CT volumes through a dense pyramid structure that integrates low-level, high-resolution features with high-level semantic features. This architecture enables the detection of both minute and large nodules, producing a comprehensive list of 3D bounding box candidates.
High Sensitivity and Specificity (HS²) Network
The core innovation is the HS² network, designed specifically for FP reduction. It introduces the Location History Image (LHI), a concept inspired by Motion History Images in video analysis. By treating the Z-axis (sequential CT slices) as a temporal dimension, the LHI projects the spatial variance of a candidate across slices into a single 2D representation.
Key Insights: The Location History Image (LHI)
The LHI provides a powerful morphological signature. While a cross-section of a blood vessel and a nodule might look identical on a single slice, their behavior across multiple slices is distinct. True nodules typically expand and contract concentrically, resulting in a circular LHI with a distinct center. In contrast, linear structures like vessels shift position across slices, creating a directional "trajectory line" in the LHI.
By explicitly encoding these geometric truths, the LHI allows a lightweight 2D CNN to effectively differentiate true nodules from lookalikes.
Results
The framework was evaluated on the LUNA16 dataset (888 CT scans, 1,186 nodules) using 10-fold cross-validation.
- State-of-the-Art Accuracy: The model achieved a Competition Performance Metric (CPM) score of 0.952.
- Superior Specificity: At a strict threshold of 1/8 false positive per scan, the model reached 90.4% sensitivity, outperforming previous state-of-the-art results by 15.6%.
- Clinical Relevance: At 2 FPs per scan, the model maintains a high sensitivity of 97.14%.
- Effective Refinement: Ablation studies showed the HS² network eliminated 84.5% of false positives generated by the initial 3DFPN stage while preserving true positives.
Conclusion
The 3DFPN-HS² framework demonstrates that specialized pipelines—combining high-recall 3D networks with domain-informed 2D refinement—can significantly improve the specificity of CAD systems. While the multi-stage approach adds computational complexity compared to single-pass models, the dramatic reduction in false positives brings automated lung cancer screening much closer to practical clinical utility.
Summary updated