Abstract
Accurately distinguishing between benign and malignant pulmonary ground-glass nodules (GGNs) remains a significant clinical challenge. This study introduces the CB-DR (Clinic-Biomarker-combined Deep Radiomic) model, a multimodal deep learning approach that integrates clinical data, blood-based biomarkers (TPI-1 and miR-206), and 3D deep radiomics. Validated through a multicenter prospective cohort, the CB-DR model exhibited satisfactory performance in predicting GGN malignancy, offering a potential tool to reduce overtreatment and guide clinical management.
Introduction
The widespread adoption of low-dose CT (LDCT) screening has led to a surge in the detection of pulmonary ground-glass nodules. While many GGNs are early-stage lung cancers, a substantial number are benign inflammatory or infectious lesions. Current diagnostic methods often struggle with the "gray area" where radiological features overlap, leading to either unnecessary invasive procedures (overtreatment) or patient anxiety and repeated radiation exposure from long-term monitoring. This research addresses the need for a non-invasive, highly accurate multimodal signature to improve diagnostic precision.
Methods and Model Architecture
The researchers developed and compared six models, ultimately focusing on the CB-DR model. This architecture utilizes Deep Radiomics (DR) based on a 3D RegNet—a convolutional neural network (CNN) optimized via neural architecture search.
To enhance spatial awareness, the model incorporates "distance maps," which are 3D representations of the signed distance from each voxel to the nodule surface. This allows the network to better interpret regional context, such as spiculation or lobulation at the nodule boundary. Furthermore, the model integrates two specific molecular biomarkers: TPI-1 (a glycolytic enzyme) and miR-206 (a microRNA), which serve as a "molecular lens" to capture metabolic changes preceding definitive radiological growth.
Technical robustness was ensured through pretraining on the LUNA16 dataset and the use of manifold mixup for data augmentation. Explainability was addressed using Class Activation Mapping (CAM) to verify that the model's predictions were driven by relevant features within the nodule and its immediate environment.
Datasets and Study Design
The study utilized a multicenter prospective cohort consisting of 501 participants (571 GGNs) recruited from seven medical centers across China between 2021 and 2023. The data was split into a training-validation group and an independent test group to ensure the generalizability of the results.
- Training-Validation Group: 407 participants (454 GGNs) from Centers 1 and 2.
- Independent Test Group: 94 participants (117 GGNs) from Centers 3, 4, 5, and 6.
Key Results and Insights
The CB-DR model demonstrated superior performance across several metrics:
- Predictive Accuracy: The model achieved an AUC of 0.90 on the independent test set, with a sensitivity of 0.90 and specificity of 0.82.
- Subgroup Performance: Performance was particularly high for mixed GGNs (mGGNs), reaching an AUC of 0.99.
- Clinical Impact: Simulation of clinical stratification suggested the model could potentially reduce overtreatment for 82.4% of benign cases.
- Calibration: A Brier score of 0.07 indicated that the model's predicted probabilities were highly reliable and closely matched actual malignancy rates.
The findings confirm that integrating clinical, molecular, and deep radiomic data provides a more comprehensive diagnostic picture than any single modality alone.
Limitations and Future Work
Despite its success, the study noted a high prevalence of malignant GGNs (~90%) in the surgical centers involved, which may not fully reflect a general screening population. Additionally, the current reliance on manual segmentation by pulmonologists presents a bottleneck for high-throughput clinical use. Future research will focus on automating segmentation and validating the CB-DR model in more diverse, global populations to confirm its efficacy across different ethnicities and healthcare systems.
Summary updated