What Are Subsolid Nodules? Subsolid nodules (SSNs) are hazy patches visible on CT scans that do not completely obscure the surrounding lung tissue. They are increasingly detected thanks to widespread low-dose CT screening and are considered strong early predictors of lung adenocarcinoma.
The Core Problem Not all SSNs grow or become malignant. Most are indolent - they sit stable for years. But some grow rapidly and require intervention. Distinguishing between these groups with conventional CT measurements (like nodule diameter) has been unreliable.
The Cost of Uncertainty Patients with ambiguous SSNs face years of repeat CT surveillance, causing anxiety, radiation exposure, and healthcare costs. A tool that accurately predicts which nodules will grow could spare low-risk patients from unnecessary surveillance while flagging high-risk ones for earlier intervention.
This Study's Approach Researchers enrolled 353 patients with 387 SSNs and developed three models - a radiomics model extracting 1454 quantitative CT features, a deep learning model using ResNet18 architecture, and a combined fusion model integrating both. All models were tested for their ability to predict which nodules would grow during follow-up.
Radiomics Feature Extraction From each nodule's 3D CT volume, 1454 raw features were extracted spanning six categories: first-order statistics (intensity distributions), gray-level co-occurrence matrix (texture), shape descriptors, and four additional spatial texture matrices. Only reproducible features with intra- and inter-observer agreement above 0.75 were kept.
Feature Selection and Model Building To reduce 1454 features to the most predictive subset, researchers used Pearson correlation filtering followed by LASSO regression, ultimately retaining 30 features. Seven machine learning algorithms were compared - including random forest, XGBoost, and logistic regression - with random forest achieving the best radiomics-only performance.
Deep Learning with ResNet18 The deep learning model used the ResNet18 architecture, a well-established 2D convolutional neural network. Images from the CT slice with the largest nodule cross-section, plus slices above and below, were used as input. Grad-CAM heatmaps highlighted which image regions most influenced predictions - providing visual interpretability.
Fusion Model The combined model concatenated deep learning features extracted from ResNet18's penultimate layer with the 30 radiomics features, then passed this joint vector through fully connected layers for final classification. The key innovation is that radiomics captures hand-crafted statistical properties while deep learning captures emergent visual patterns - together they provide more information than either alone.
Model Comparison In the test set of 78 nodules, the combined model achieved an AUC of 0.926, the radiomics-only model achieved 0.894, and the deep learning-only model achieved 0.802. The combined model's superiority over the deep learning model was statistically significant (DeLong test p = 0.012).
Balanced Sensitivity and Specificity The combined model achieved 84.6% sensitivity and 89.7% specificity - meaning it correctly identified 85% of growing nodules while correctly clearing 90% of stable ones. This balance is clinically valuable as it avoids both missed cancers and unnecessary interventions.
Why Radiomics Outperformed Pure Deep Learning The dataset of 387 nodules is relatively small for training deep learning from scratch. Radiomics features, being hand-engineered from domain knowledge, carry strong signal even with limited data. Deep learning requires more data to generalize well, which is why the fusion approach gains from radiomics' efficiency.
Calibration Results The Hosmer-Lemeshow calibration test showed the combined model (p = 0.308) and radiomics model (p = 0.927) were well-calibrated - meaning their predicted probabilities match observed outcomes. The deep learning model showed poor calibration (p = 0.003), another reason fusion improves upon standalone deep learning.
Risk Stratification at Diagnosis When a patient is found to have a subsolid nodule on a screening CT, this model could immediately classify it as low or high risk for growth. Low-risk patients could be monitored less intensively, reducing radiation and anxiety. High-risk patients could receive closer follow-up or earlier intervention.
Nodule Characteristics That Predict Growth Growing nodules tended to be larger (mean diameter 13.5 mm vs. 7.8 mm), older patients, and more often of the part-solid type rather than pure ground glass. The radiomics model confirmed that textural features beyond visible diameter significantly improve prediction.
Grad-CAM for Trust Building The Grad-CAM activation maps showed which regions of the CT image the deep learning model was using. When these maps focus on the nodule itself rather than surrounding tissue, clinicians can trust the model's reasoning. This kind of visual explanation is essential for clinical adoption.
Comparison to Prior Work Previous radiomics models for SSN growth prediction achieved AUCs of 0.780 to 0.911 in smaller studies. This study's combined model at AUC 0.926 with a relatively larger and more diverse sample is competitive, and the fusion approach adds a methodological advance that others can build upon.
Single-Institution Retrospective Data All patients came from one Chinese hospital, limiting generalizability. Variations in CT scanner models, imaging protocols, and population demographics could affect performance when the model is applied elsewhere.
Manual Segmentation Dependency Nodule boundaries were manually delineated by radiologists, introducing potential variability and limiting scalability. Automating segmentation through AI-based tools would be necessary for deployment in high-volume screening programs.
Short Follow-Up for Some Nodules Nodules were defined as non-growing if they remained stable across follow-up, but some may eventually grow beyond the study's observation window. Longer follow-up periods would improve outcome accuracy.
Future Directions Incorporating clinical data (smoking history, family history, lung cancer risk scores) alongside imaging features could further improve the model. Prospective multicenter validation and integration with radiologist workflow tools are needed before clinical deployment.