Lung Cancer's Devastating Toll. Lung cancer is the most common cause of cancer deaths worldwide, with only an 18% five-year survival rate. The low survival rate stems largely from late detection - by the time symptoms appear, cancer has often spread beyond the lungs, making treatment far more difficult and less effective.
CT Imaging as the Diagnostic Standard. Computed tomography (CT) scanning is the preferred diagnostic method because it provides high-resolution cross-sectional views of the lungs from multiple angles, enabling three-dimensional evaluation of suspicious lesions. CT can detect small nodules that might be missed by X-rays, which provide less detail about soft tissues.
The Problem with Manual Analysis. Manual inspection of CT scans by radiologists is time-consuming, prone to errors, and subject to inter-observer variability. Early-stage nodules can closely resemble surrounding normal structures on CT scans, making accurate early identification particularly challenging even for experienced clinicians.
The Opportunity for AI. Automated AI-based systems offer the potential to overcome these limitations by consistently analyzing CT images without fatigue, reducing errors, and flagging suspicious nodules that might be overlooked. The authors argue that effective feature extraction is the key to improving AI accuracy in this domain.
Combining Traditional and Deep Learning Features. The core innovation of this study is the systematic combination of traditional image analysis methods (GLCM and SIFT) with deep learning architectures (VGG-16 and MobileNet) to extract complementary types of information from CT images. The hypothesis is that no single approach captures all discriminative information, so combining them produces a richer and more reliable feature set.
Gray-Level Co-occurrence Matrix (GLCM). GLCM is a classical texture analysis technique that captures how pairs of pixels with specific gray values are spatially related throughout an image. From the co-occurrence matrix, the researchers extract nine texture features including energy, contrast, correlation, homogeneity, variance, and entropy. These features are particularly useful for distinguishing the fine texture differences between smooth benign nodules and irregular malignant ones.
Scale-Invariant Feature Transform (SIFT). SIFT detects stable keypoints in an image and describes them with 128-dimensional vectors that remain consistent regardless of image rotation, scaling, or lighting changes. For CT images, which may be acquired at different orientations and with different scanner settings, SIFT's invariance properties help capture structural edges and geometric contours reliably across varying conditions.
Six Tested Feature Combinations. Rather than simply proposing one fusion approach, the study systematically evaluated six combinations: GLCM+MobileNet, GLCM+VGG-16, SIFT+MobileNet, SIFT+VGG-16, GLCM+SIFT+MobileNet, and GLCM+SIFT+VGG-16. This systematic comparison allows rigorous identification of which combination offers the best balance of accuracy, computational cost, and reliability.
VGG-16: Deep and Hierarchical. VGG-16 is a 16-layer neural network with 13 convolutional layers arranged in five blocks, followed by three fully connected layers. The network uses small 3x3 convolution kernels throughout, which collectively capture fine image details at multiple scales. As the network deepens, the number of filter channels increases from 64 to 512, enabling progressively more abstract and complex feature representations.
MobileNet: Lightweight and Efficient. MobileNet was designed for deployment on devices with limited computing power, using a novel technique called depthwise separable convolution. Instead of applying a standard convolution (which combines spatial and channel information simultaneously), MobileNet first applies convolution separately to each channel, then combines channels with a 1x1 pointwise convolution. This dramatically reduces the number of parameters and computational cost while preserving accuracy.
Why These Two Architectures Were Chosen. VGG-16 and MobileNet were selected because they offer complementary strengths: VGG-16 provides stable, interpretable hierarchical representations well-suited for moderate-sized datasets, while MobileNet enables fast inference on CPU hardware. Both are simpler than more complex architectures like ResNet or EfficientNet, making them appropriate for the moderate dataset size in this study.
Feature Extraction from Deep Networks. For both architectures, features are extracted not from the final classification layer but from the penultimate layer. This captures rich semantic representations learned from the visual patterns in CT images while avoiding the architecture-specific classification decisions of the final layer. These deep features are then fused with the handcrafted GLCM and SIFT features.
IQ-OTH/NCCD Dataset. The study uses a publicly available Iraqi lung cancer dataset of 1,097 CT scans from 110 patients collected in 2019 at the Iraq-Oncology Teaching Hospital and National Center for Cancer Diseases. The images were labeled by oncologists and radiologists as normal, benign, or malignant. Because the benign class was underrepresented, data augmentation was used to balance the dataset to 1,000 images per class (3,000 total).
Preprocessing Pipeline. All images were resized to 224x224 pixels to standardize input dimensions. A Gaussian filter was applied to reduce CT noise arising from patient size, scanner settings, or radiation exposure. Data augmentation was applied to training images only, including zoom (factor 0.3), shear (0.3), and vertical flipping to increase diversity and prevent overfitting.
Feature Normalization and Fusion. GLCM features were computed using horizontal orientation at distance 1, producing a 20-dimensional texture vector. SIFT features were encoded using a bag-of-visual-words approach with 256 visual words. All three feature blocks were independently standardized using training set statistics, then concatenated and reduced to 256 dimensions using Principal Component Analysis (PCA) to remove redundancy before classification.
Classification Head. The fused and dimensionality-reduced features were passed through a fully connected layer consisting of a Dense(32) layer with ReLU activation, a Dropout(0.3) layer for regularization, and a final Dense(3) layer with Softmax activation for three-class output (normal, benign, malignant). Training used cross-entropy loss with SGD optimization for 100 epochs with early stopping based on validation loss.
Best Overall: GLCM + SIFT + MobileNet. The winning combination across all metrics was the fusion of all three feature types - GLCM texture, SIFT geometry, and MobileNet deep features. This combination correctly classified 298 out of 300 test images, achieving an overall accuracy of 99.33% on test data and 98.33% when averaged comprehensively, with precision of 98.34%, recall of 98.34%, F1-score of 98%, and specificity of 98.67%.
MobileNet Consistently Outperforms VGG-16. Across all pairing comparisons (GLCM+MobileNet vs. GLCM+VGG-16, SIFT+MobileNet vs. SIFT+VGG-16, and GLCM+SIFT+MobileNet vs. GLCM+SIFT+VGG-16), the MobileNet-based models achieved higher accuracy. The GLCM+SIFT+VGG-16 combination achieved 98% accuracy, a strong result but still trailing the MobileNet version.
Comparison with Prior Methods. When compared with previous studies using the same IQ-OTH/NCCD dataset, the proposed model achieves 98.33% accuracy with comprehensive metrics reported (precision, recall, F1, specificity). Prior studies reached 93-97% accuracy but often evaluated only some metrics. The proposed framework consistently outperforms existing approaches across all reported measures.
Why the Triple Fusion Works. GLCM captures fine-grained texture heterogeneity of pulmonary tissue, making it sensitive to irregular malignant patterns. SIFT encodes structural edges and geometric contours stable across CT orientations. MobileNet adds high-level contextual and semantic abstractions. Together, these form a multi-scale representation that unites micro-texture with macro-context, enabling the model to distinguish subtle differences between cancer types.
Practical Hardware Requirements. The entire system runs on a standard desktop CPU (Intel Core i5-9500T with 8 GB RAM), without requiring specialized GPU hardware. The proposed GLCM+SIFT+MobileNet combination processes each CT image in an average of 0.82 seconds with a peak memory usage of approximately 1.6 GB, making it feasible for clinical decision support in standard office or clinic computing environments.
Throughput Analysis. The proposed hybrid system achieves a throughput of 73 images per minute - lower than standalone MobileNet (98/min) due to the added feature extraction steps, but still fast enough for practical use. The trade-off in speed is justified by the significant accuracy gain: pure MobileNet alone achieves lower accuracy than the full hybrid combination.
External Validation. To test generalizability, the trained model was applied without modification to 10 CT images from an independent Kaggle lung CT dataset. It achieved 97.2% accuracy and an F1-score of 0.96 on this external data without any retraining, suggesting the learned features transfer across datasets. The researchers acknowledge this validation set is small and plan larger multi-institutional validation as future work.
Path to Real-Time Deployment. The current system is suitable for desktop-based clinical decision support. Future planned work includes model pruning (removing less important parameters), quantization (reducing numerical precision), and lightweight fusion strategies to enable real-time inference on portable or mobile platforms, expanding access to low-resource healthcare settings.
Single Dataset Limitation. The study relies primarily on one dataset (IQ-OTH/NCCD), which limits the conclusions that can be drawn about generalizability to other patient populations, scanner types, or geographic settings. The authors acknowledge this as the primary limitation and commit to validating the model on additional publicly available datasets in future work.
Dataset Size Constraints. With 1,097 original CT scans from 110 patients, the primary dataset is modest by deep learning standards. Heavy data augmentation was used to compensate, expanding the effective training set, but a larger and more diverse dataset would provide stronger evidence of model robustness and real-world applicability.
Three-Class Clinical Value. The ability to classify CT scans into three categories - normal, benign, and malignant - is clinically meaningful because it directly supports the clinical decision of whether to monitor, investigate further, or immediately treat a suspicious lung nodule. This three-way distinction is more useful than binary cancer/non-cancer classification for guiding patient management.
Broader Clinical Integration. As AI tools for lung cancer detection advance, their integration into clinical workflows will require not just high accuracy but also interpretability, regulatory approval, and prospective clinical validation. The authors' open-source code availability and plan for multi-institutional validation are important steps toward eventual clinical translation of hybrid approaches like this one.