Clinical Problem: In lung cancer screening with LDCT, newly detected intermediate-sized nodules (50-500 mm3) require follow-up CT scans, but about 55% of new nodules eventually disappear on their own. Following up every new nodule generates unnecessary radiation, cost, and patient anxiety.
Study Goal: This study developed a multi-view convolutional neural network (CNN) to predict which newly detected intermediate-sized solid nodules will resolve (disappear) on follow-up, enabling safe reduction of unnecessary follow-up CTs.
Data Source: 344 intermediate-sized nodules from 250 participants in the NELSON (Dutch-Belgian Randomized Lung Cancer Screening) trial - one of the world's most rigorous and well-characterized lung cancer screening datasets.
Key Performance: The multi-view CNN achieved AUC 0.81, mean sensitivity 0.63, and mean specificity 0.93, significantly outperforming all five comparison models (2D, 2.5D, and 3D single-view models, p < 0.05).
NELSON Screening Context: The NELSON trial demonstrated 26% lung cancer mortality reduction with LDCT screening. During incidence rounds, 5-7% of screening participants have newly detected nodules, requiring management decisions.
Resolving Nodule Problem: About 55% of new nodules in NELSON disappeared at subsequent scans, while only 4% were lung cancer. Following up all new nodules thus generates a 55% false-positive follow-up rate for the resolving category.
Current Management: Management is primarily based on nodule size and appearance per guidelines (Lung-RADS, NELSON protocol). There is no currently validated tool to predict whether a specific new nodule will resolve.
Clinical Impact: If non-resolving nodules requiring follow-up can be reliably identified at specificity >90% (missing <10% of non-resolving nodules), follow-up CT could be safely eliminated for predicted-resolving nodules, reducing radiation and healthcare costs.
Multi-View Design: Three 2D ResNet-18 models process axial, coronal, and sagittal middle slices respectively (each using 3 consecutive slices stacked as channels). One 3D ResNet-18 model processes the full 32x32x32 mm3 cubic nodule volume.
Feature Fusion: The four networks' outputs are concatenated into a 72-dimensional feature vector, which is passed through a multi-layer perceptron to produce final class probabilities. Late fusion combines complementary 2D and 3D information.
Preprocessing: Lung window standardization (WW 1600 HU, WL -700 HU), B-spline interpolation to 1x1x1 mm isotropic voxels, and nodule centroid-centered cropping ensure consistent input quality across different scanners and CT protocols.
Explainability: Grad-CAM++ was applied to the multi-view model, generating heatmaps that highlight the most influential image regions for resolving versus non-resolving nodule predictions, aiding radiologist interpretation.
Multi-View AUC: The multi-view model achieved AUC 0.81 (95% CI: 0.75-0.88), significantly outperforming all five comparison models: 2D axial (0.68), 2D coronal (0.68), 2D sagittal (0.67), 2.5D (0.67), and 3D (0.71), all p < 0.05.
Sensitivity and Specificity: At the operating point maximizing specificity, the multi-view model achieved mean sensitivity 0.63 (SD 0.15) and mean specificity 0.93 (SD 0.02), demonstrating reliable identification of non-resolving nodules.
Clinical Benefit Calculation: With specificity greater than 90% (meaning fewer than 10% of non-resolving nodules are missed), follow-up CT could be prevented in 14% of participants who would otherwise require it.
Sub-Analysis: Excluding 10 'nearly-resolved' nodules (in process of resolving but not yet gone) in a sensitivity analysis did not substantially change model performance, confirming robustness of the main findings.
Grad-CAM++ Heatmaps: The Grad-CAM++ algorithm visualized which image regions most influenced the model's resolving vs. non-resolving predictions, generating spatial attention maps overlaid on CT slices.
Model Focus: Heatmaps revealed that the model attended to both the nodule core and its surrounding parenchyma, suggesting that peri-nodular CT texture features contribute to resolving prediction alongside the nodule's own appearance.
Radiologist Validation: Examining model-highlighted regions alongside known imaging features of resolving nodules (e.g., ground-glass component, indistinct margins) can validate that the CNN's learned features align with clinical reasoning.
Trust-Building Role: Explainable AI heatmaps are critical for clinician acceptance; radiologists are more likely to act on AI predictions when they can see that the model focuses on clinically relevant image features.
14% Scan Reduction: If 14% of follow-up CTs can be safely eliminated (by confidently predicting which nodules will resolve), this translates to significant reductions in cumulative radiation exposure, healthcare costs, and patient anxiety across screening programs.
Screening Algorithm Integration: The multi-view model could be integrated as an additional decision layer in existing LDCT screening algorithms, providing a probability score for nodule resolution that complements current size-and-appearance based follow-up recommendations.
Workflow Design: AI-predicted resolving nodules could be flagged for radiologist review with a suggested shorter surveillance interval rather than immediate follow-up CT, maintaining clinical oversight while reducing unnecessary imaging.
Patient Communication: Providing patients with a quantified probability that their newly detected nodule will resolve could reduce anxiety compared to being told simply that a follow-up scan is needed in 3-6 months.
Single Screening Trial Data: Data exclusively from the NELSON trial (Dutch-Belgian, predominantly male, heavy smokers, 2004-2006 vintage scans) may not generalize to modern screening programs, different demographics, or current scanner technology.
Moderate Sensitivity: Sensitivity of 0.63 means 37% of truly resolving nodules are not identified as such by the model, limiting the proportion of follow-up CTs that could be safely eliminated. Improving sensitivity without sacrificing specificity remains a key challenge.
Nodule Annotation Variability: Approximate centroid annotation (rather than full volumetric segmentation) may miss edge-located features. Full automated segmentation could improve input quality and model performance.
External Prospective Validation: Before clinical deployment, the model must be validated in modern LDCT screening cohorts from multiple programs, scanners, and countries, with prospective follow-up confirming that predicted-resolving nodules safely disappear without missed cancers.