Non-small cell lung cancer accounts for about 85% of all lung cancer cases, and its two main subtypes require fundamentally different treatments. Adenocarcinoma (ADC), representing about 45% of cases, is often treated with targeted therapies aimed at mutations like EGFR. Squamous cell carcinoma (SCC), representing about 35%, typically receives platinum-based chemotherapy or immune checkpoint inhibitors. The wrong treatment for the wrong subtype can be harmful or ineffective.
Currently, NSCLC subtype identification depends on pathological analysis of biopsy tissue, which has significant limitations. Obtaining a biopsy is invasive and painful, and some patients are too ill to safely undergo the procedure. Even when biopsies are obtained, only a small portion of the tumor is sampled, which may not capture the full biological heterogeneity of the lesion.
CT imaging offers a non-invasive window into tumor characteristics that could replace or supplement biopsy for subtype classification. Both radiomics (quantitative features extracted from images) and deep learning (features learned automatically by neural networks) have shown promise individually. This study tested whether combining both approaches would outperform either one alone.
The hybrid approach aims to leverage the complementary strengths of handcrafted radiomics features and automatically learned deep features. Radiomics captures interpretable tumor properties like shape, texture, and density. Deep learning captures abstract patterns invisible to the human eye. Together they may provide a more complete characterization of each tumor than either approach in isolation.
The study analyzed CT images from 235 patients with NSCLC, balanced between 110 adenocarcinoma and 112 squamous cell carcinoma cases. All CT scans were obtained with contrast enhancement using a standardized protocol. The dataset was split 75% for training and 25% for testing, with an additional 48-patient external validation set drawn from the publicly available NSCLC-Radiomics database on the Cancer Imaging Archive.
A total of 1,409 radiomics features were extracted from three-dimensional tumor volumes using the PyRadiomics package. These included shape, intensity, and texture features from original CT images, as well as intensity and texture features derived from 14 image filters including wavelet transformations and gradient filters. This comprehensive extraction ensured that a wide range of quantifiable tumor characteristics was captured.
Deep features were extracted using VGG-19, a pre-trained convolutional neural network originally trained on the ImageNet database. For each patient, the CT slice containing the largest cross-sectional tumor area was selected, and the tumor region was cropped and zero-padded outside the region of interest. The model was fine-tuned using stochastic gradient descent, and 8,192 features were extracted from the two fully connected layers.
Both feature sets underwent dimensionality reduction via PCA followed by feature selection using L2,1-norm minimization. PCA reduced the 1,409 radiomics features and 8,192 deep features to 100 components each, which were then concatenated into 200 combination features. The final selected feature sets consisted of 27 radiomics features, 20 deep features, and 55 combined features (25 radiomics plus 30 deep). Classification used an ensemble of bagging, AdaBoost, and logistic regression with soft voting.
The combined radiomics and deep learning model achieved the highest AUC and accuracy on both internal and external test sets. On the internal test set of 55 patients, the hybrid model reached an AUC of 0.786 and accuracy of 0.691, compared to 0.643 and 0.614 for radiomics alone, and 0.704 and 0.655 for deep features alone.
The performance advantage held on the external validation set drawn from a different institution and imaging system. The hybrid model achieved an AUC of 0.720 and accuracy of 0.646 on the external NSCLC-Radiomics dataset, compared to 0.611 and 0.583 for radiomics alone, and 0.662 and 0.625 for deep features alone. This consistency across different data sources demonstrates the model's generalizability.
Compared to the radiomics-only model, the hybrid approach improved AUC by 22% on the internal set and 18% on the external set. Accuracy improved by 13% and 11%, respectively. The relative improvement in AUC exceeded the improvement in accuracy across both datasets, suggesting the combined model produces more balanced and robust predictions especially on unequal data distributions.
Among the top 20 most important features in the combined model, deep features dominated, contributing 60.85% of the total importance. The top-ranked deep component achieved an importance score of 0.162, outperforming the highest-ranked radiomics feature at 0.110. This suggests that the abstract representations captured by deep learning add complementary information beyond what handcrafted radiomics features alone can capture.
Radiomics features excel at characterizing measurable tumor properties like shape irregularity, texture roughness, and intensity distribution that differ between adenocarcinoma and squamous cell carcinoma. However, radiomics-only models showed a consistent bias toward predicting SCC, likely because SCC has more prominent and easily quantifiable structural characteristics that dominate the feature space.
Deep features corrected this bias by providing high-level abstract representations that capture patterns not quantifiable by traditional radiomics. When deep features were added, the model's ability to correctly identify adenocarcinoma improved substantially, and overall predictions became more balanced between the two subtypes. This complementarity is the key advantage of the hybrid approach.
The hybrid model outperformed both prior radiomics-based studies (AUC 0.700) and deep learning-based studies (AUC 0.710) from the recent literature. By integrating both feature types, the model effectively combines interpretable quantitative tumor descriptors with the pattern recognition power of deep networks, exceeding what either technology could achieve independently.
Validation on an external dataset from a different institution, scanner, and patient demographics demonstrates the model's ability to generalize beyond its training environment. This is a critical requirement for any clinical tool, as real-world deployment involves diverse patients and imaging equipment. Many prior studies were evaluated only on single-institution datasets, limiting confidence in their broader applicability.
Starting with 1,409 radiomics and 8,192 deep features, the pipeline needed to reduce dimensionality dramatically to build a practical classifier. High-dimensional feature spaces suffer from the curse of dimensionality, increased risk of overfitting, and high computational costs. The two-step reduction process using PCA followed by L2,1-norm minimization addressed these challenges systematically.
PCA reduced each feature set to 100 principal components while preserving over 90% of the variance in the original data. For deep features, 88 principal components explained 90% of variance, and for radiomics features 56 components sufficed. Retaining 100 components in each case preserved additional potentially useful variance while achieving a major reduction in dimensionality from thousands to hundreds of features.
L2,1-norm minimization then selected the most informative features from the combined 200-component pool. This mathematical approach computes the L2-norm for each row of the feature matrix and then applies L1-norm to enforce sparsity, effectively zeroing out uninformative features while preserving the most discriminative ones. The final selections of 27 radiomics, 20 deep, and 55 combination features represent a roughly 95% reduction from the original feature spaces.
The ensemble classifier combined bagging, AdaBoost, and logistic regression using soft voting to balance variance reduction, bias minimization, and interpretability. Bagging trains on random data subsets to reduce variance, AdaBoost focuses on misclassified samples to reduce bias, and logistic regression provides transparent decision boundaries important for clinical applications. The soft voting mechanism combines the probability outputs of all three, rather than using simple majority voting.
The hybrid model offers a non-invasive alternative or complement to biopsy for NSCLC subtype classification that could benefit patients unsuitable for surgical procedures. Patients with advanced disease, poor performance status, or tumor locations that make biopsy technically difficult or dangerous could benefit most. CT-based classification requires only imaging that is already obtained during standard diagnostic workup.
The model's ability to distinguish between adenocarcinoma and squamous cell carcinoma has direct treatment implications. Correct subtype identification guides whether a patient receives EGFR-targeted therapy, platinum-based chemotherapy, or immunotherapy. Misclassification could result in a patient receiving an ineffective or potentially harmful treatment regimen, making accuracy critical even when performance is not perfect.
Future integration of PET imaging data alongside CT could further improve classification performance. Prior studies have shown that NSCLC subtypes have distinct metabolic profiles visible on PET scans. Combining the spatial tumor characterization from CT radiomics and deep features with the metabolic information from PET could create a more complete multi-modal portrait of each tumor.
Explainable AI methods and model simplification are needed before this approach can be integrated into routine clinical workflows. Currently the model functions as a black box for the deep feature components, making it difficult for clinicians to understand why a particular classification was made. Incorporating tools that highlight which imaging regions and features drove each prediction would build clinician trust and support regulatory approval.
This study demonstrates that combining radiomics and deep learning features from CT images produces a more accurate NSCLC subtype classifier than either approach used alone. The hybrid model achieved an AUC of 0.786 on internal validation and 0.720 on external validation, outperforming prior single-modality approaches from the literature.
The complementarity of the two feature types is the central finding: radiomics captures what can be measured, and deep learning captures what cannot be easily measured. Together they provide a richer representation of each tumor, reducing the prediction bias inherent in radiomics-only models and combining interpretable measurements with powerful pattern recognition.
The limitations of the study include its relatively small sample size of 235 patients from a single institution, restriction to only two NSCLC subtypes, and reliance on a single CT scanner and protocol. Future work should expand to multi-center datasets, incorporate additional subtypes, explore more diverse deep learning architectures beyond VGG-19, and investigate multi-modality approaches combining CT and PET.
With further validation and clinical integration, this hybrid approach has the potential to become a reliable non-invasive tool for guiding treatment decisions in NSCLC. As CT imaging is already part of standard diagnostic workup for lung cancer, adding automated hybrid classification would not require additional procedures, making this a practically accessible innovation for improving personalized treatment selection.