Lung cancer remains the leading cause of cancer mortality worldwide, and early detection is critical. Approximately 237,000 new lung cancer cases were diagnosed in the United States in 2023 alone, with the high fatality rate directly tied to advanced-stage presentation. CT imaging has become the standard modality for early detection, but manual interpretation is time-consuming and subject to variability between radiologists, creating an urgent need for automated diagnostic assistance.
Prior AI approaches to lung cancer detection relied on either deep learning or traditional machine learning, each with fundamental limitations. CNN-based approaches like VGGNet-16 excel at automatically extracting hierarchical features from images but are prone to overfitting on small medical datasets without strong regularization. Traditional machine learning classifiers like SVM offer robust classification in high-dimensional spaces with resistance to overfitting, but depend on manually engineered features that may miss subtle imaging patterns.
Lung cancer manifests in three histologically distinct subtypes that require different clinical management. Adenocarcinoma originates in glandular cells on the lung's outer surface and is the most common subtype, particularly in non-smokers. Large cell carcinoma is characterized by rapid growth and can arise anywhere in the lung. Squamous cell carcinoma is closely associated with smoking and typically originates in the central bronchial region. Distinguishing all three from normal tissue simultaneously is a more clinically relevant and challenging task than binary cancerous versus non-cancerous classification.
Previous studies identified six specific limitations that SVMVGGNet-16 was designed to address. These limitations were: restricted multi-class classification relying on binary outputs; dependence on a single model architecture; inadequate image pre-processing techniques; use of singular performance metrics like accuracy alone; limited clinical applicability due to poor generalizability; and challenges with overfitting on limited datasets. Each limitation had direct consequences for diagnostic reliability in clinical deployment.
The LIDC-IDRI dataset was used due to its size, annotation quality, and public availability. The dataset contains CT scan images of three lung cancer types (adenocarcinoma, large cell carcinoma, and squamous cell carcinoma) along with normal lung tissue, organized into training, validation, and testing splits at 70, 20, and 10 percent respectively. The dataset is annotated by multiple radiologists, providing robust ground truth for model training and evaluation.
Median filtering and histogram equalization were applied to standardize and enhance image quality. The median filter reduces salt-and-pepper noise by replacing each pixel value with the median intensity value of its neighborhood, preserving edges critical for tumor boundary detection. Histogram equalization redistributes pixel intensity values to improve contrast, making subtle tumor features more visible and improving the robustness of downstream feature extraction.
Tumor segmentation used intensity thresholding combined with Canny edge detection. Thresholding separated tumor regions from surrounding tissue by classifying pixels above a defined intensity value as tumor and pixels below it as background. The Canny edge detector then precisely delineated tumor boundaries through a multi-step process: noise reduction, gradient computation, non-maximum suppression, and hysteresis-based edge tracking. This combination produced well-defined tumor region masks for geometric feature extraction.
Five geometric features were extracted from segmented tumor regions for SVM input. Area quantified tumor size by measuring the number of pixels in the segmented region. Perimeter characterized boundary shape. Eccentricity measured tumor elongation. Compactness indicated how closely the tumor approximated a circular form. Circularity provided a normalized shape descriptor. These geometric features complement the deep features extracted by VGGNet-16, providing the SVM with structured, interpretable inputs for classification.
VGGNet-16 was selected over alternative architectures due to its feature extraction depth and transfer learning capability. Compared to VGG-19, AlexNet, and GoogLeNet, VGGNet-16 offers a balance between depth and computational efficiency, with strong pre-trained representations from ImageNet that transfer effectively to medical imaging tasks. Its hierarchical convolutional layers systematically extract features at increasing levels of abstraction, from basic edges to complex tissue morphology patterns specific to ADC, LCC, SCC, and normal lung tissue.
The VGGNet-16 architecture was customized with additional layers for four-class lung cancer classification. The pre-trained VGGNet-16 base produces 7x7x512 feature maps from 224x224 pixel RGB input images. Batch normalization standardizes activations for training stability, followed by max pooling to reduce spatial dimensions to 3x3x512. Features are flattened to a 4608-dimensional vector and passed through dense layers of 1024, 512, 256, and 128 neurons with ReLU activation and dropout regularization, culminating in a softmax output layer for four-class prediction.
SVM was chosen for its resistance to overfitting and effectiveness in high-dimensional classification. SVM's ability to find optimal decision boundaries in high-dimensional feature spaces makes it particularly suited to classifying geometric features extracted from tumor images. Unlike neural networks that can memorize training data, SVM's margin-maximization objective intrinsically regularizes the model, producing reliable classification even when training examples are limited. Random Forest and AdaBoost were evaluated and found inferior to SVM for this task.
The hybrid SVMVGGNet-16 model integrates VGGNet-16 feature extraction with SVM classification. VGGNet-16 operates as the feature extractor, learning complex hierarchical representations from CT scan images during training. The features generated by VGGNet-16 are then passed to the SVM, which applies its discriminative classification capabilities to separate the four cancer classes. This two-stage integration combines the representation learning power of deep learning with the generalization robustness of kernel-based classification, addressing the core limitations of each model used alone.
SVMVGGNet-16 outperformed both VGGNet-16 and SVM alone across all evaluation phases. During training, SVMVGGNet-16 achieved 97.22% accuracy and 99.42% AUC, compared to VGGNet-16 at 90.70% accuracy and 97.78% AUC, and standalone SVM at 94.45% accuracy and 99.42% AUC. The hybrid model's ability to match SVM's AUC while substantially improving accuracy demonstrates that the combination captures complementary strengths from both constituent architectures.
Validation phase results confirmed the hybrid model's superiority with 94.72% accuracy and 97.87% AUC. VGGNet-16 achieved 83.33% validation accuracy and 92.85% AUC, while standalone SVM achieved 81.94% accuracy and 94.05% AUC. The hybrid model improved validation accuracy by 11 percentage points over VGGNet-16 and 12 percentage points over SVM, demonstrating that the integration of geometric and deep features provides substantially more robust generalization than either model individually.
In the critical testing phase, SVMVGGNet-16 achieved 96.72% accuracy and 96.87% AUC. This represents a 12.59 percentage point improvement over VGGNet-16 (84.13% accuracy) and a 12.28 percentage point improvement over standalone SVM (84.44% accuracy) in testing accuracy. The preservation of high accuracy from training (97.22%) through testing (96.72%) demonstrates minimal overfitting and strong generalization to unseen data, a key requirement for clinical deployment.
Per-class testing performance showed the highest accuracy for the normal class and lowest for squamous cell carcinoma. Normal lung tissue achieved 98.46% accuracy in testing with 100% precision, reflecting the model's confidence in identifying healthy tissue. ADC achieved 96.40% testing accuracy, LCC 97.56%, and SCC 94.48%. The relatively lower SCC performance reflects the greater morphological variability in squamous cell carcinoma presentation across different tumor locations and stages.
SVMVGGNet-16 outperforms all prior models tested on the LIDC-IDRI dataset. The historical progression of LIDC-IDRI classification accuracy shows ResNet-based approaches achieving 89.9% in 2017, Deep Autoencoder models reaching 91.2% in 2019, and MV-KBC Deep Neural Network achieving 91.6% in 2022. SVMVGGNet-16 achieves 96.72% testing accuracy, representing a 5.12 percentage point improvement over the previous best and establishing a new benchmark for multi-class lung cancer classification on this widely-used dataset.
The margin of improvement over the previous best model is largest in accuracy, with competitive AUC and recall performance. The MV-KBC model used multi-view knowledge-based collaborative deep networks specifically designed for nodule classification, yet still underperforms the simpler hybrid approach of SVMVGGNet-16 by over 5 percentage points in accuracy. This demonstrates that hybrid integration of established architectures can outperform more complex novel architectures when each component's role is carefully designed.
VGGNet-16 alone showed better generalization ability than SVM alone across testing scenarios. While SVM achieved higher training accuracy and AUC than VGGNet-16 in isolation, VGGNet-16 maintained more consistent performance from validation to testing phases. SVM showed limitations in AUC during validation and testing compared to training, suggesting that geometric features alone may not provide sufficient discriminative power for the full complexity of multi-class lung cancer subtype differentiation without the complementary deep features from VGGNet-16.
The study's primary limitation is reliance on a single dataset from a single imaging modality. All experiments used CT images from the LIDC-IDRI dataset, which may limit generalizability to patients imaged with different protocols, scanner types, or in healthcare settings with different population characteristics. The absence of MRI or PET scan data also means the model cannot leverage complementary anatomical or metabolic information that these modalities provide for comprehensive diagnosis.
Retrospective study design limits direct validation of real-world clinical utility. Performance metrics were calculated on held-out test data from the same dataset distribution as training data, not from prospective clinical deployment. Real-world utility requires validation on prospective patient cohorts, clinical trials, and integration with actual radiologist workflows to identify failure modes that laboratory evaluation cannot detect. The study acknowledges this limitation and identifies prospective clinical trials as a key priority for future work.
The SVMVGGNet-16 model has specific clinical applications as a radiologist decision-support tool. By automating the initial classification of CT scan images into ADC, LCC, SCC, and normal categories, the model can reduce radiologist workload and diagnostic errors in routine screening workflows. Integration into oncology screening centers and hospital imaging infrastructure could shorten the time from scan acquisition to preliminary classification, enabling faster referral and treatment initiation for patients with early-stage malignancy.
The hybrid model could serve as an educational tool bridging data science and medical practice. The combination of VGGNet-16 and SVM illustrates how complementary computational methodologies can be integrated to address clinical diagnostic challenges, providing an instructive framework for medical trainees and practitioners learning to evaluate AI-assisted diagnostic systems. The model's interpretable geometric feature extraction provides a mechanism for explaining classification decisions in terms clinically meaningful to radiologists.
SVMVGGNet-16 demonstrates the potential of hybrid deep learning and machine learning for multi-class lung cancer classification. The model achieved 97.22% training accuracy and 96.72% testing accuracy on the LIDC-IDRI dataset, outperforming all prior benchmarks on this dataset. The combination of VGGNet-16's deep feature extraction with SVM's discriminative classification capability addresses the core limitations of each approach used independently, producing a more robust and generalizable diagnostic model.
Future work should focus on expanding dataset diversity and exploring additional architectural combinations. Priority research directions include training on multiple datasets from different imaging centers and populations to improve generalizability, exploring alternative deep learning architectures beyond VGGNet-16, and investigating advanced SVM kernel functions that may better capture the geometric complexity of lung tumor shapes. Augmenting the training data with synthetic samples from varied imaging conditions could also improve model robustness to real-world scanner variation.
Integration of explainability techniques is identified as a critical step toward clinical adoption. The current model provides classification outputs without interpretable explanations of which image features drove each decision. Incorporating gradient-based visualization methods or attention mechanisms that highlight diagnostically relevant image regions would allow radiologists to verify that the model focuses on clinically meaningful features rather than spurious patterns, increasing trust and enabling informed correction when the model errs.