A global burden with limited tools. Lung cancer accounts for 2.5 million new cases annually and 18.7% of all cancer deaths worldwide. Despite this scale, rural and resource-limited settings often lack access to the CT scanners that are standard for diagnosis.
Why chest X-rays matter. Chest X-rays (CXR) are far more affordable and widely available than CT scans. However, their lower image quality and higher noise levels have historically made automated cancer detection from CXR a challenging problem.
The promise of AI-assisted screening. Computer-aided diagnosis (CAD) systems powered by convolutional neural networks (CNNs) can process CXR images to detect cancerous patterns, potentially enabling early detection in settings where CT is unavailable. Automatic diagnosis systems improve early detection and can boost treatment success rates by up to 20%.
The small dataset problem. The JSRT dataset used for training contains only 247 chest X-ray images, 154 with lung cancer and 93 without. Training a deep CNN on such a small dataset typically leads to overfitting, where the model memorizes training examples rather than learning generalizable patterns.
Dividing images into overlapping patches. To multiply the effective training data, each CXR image was divided into a set of overlapping patches rather than using the whole image at once. Overlapping patches capture contextual information from neighboring regions, helping the model detect subtle localized abnormalities like small nodules.
Finding the optimal patch count. Experiments tested configurations of 8, 11, and 15 patches per image. Using 8 patches gave 82.47% accuracy but missed parts of the image. Using 15 patches added computational overhead without improving accuracy. Eleven patches achieved 83.19% accuracy and was selected as the best compromise between coverage and efficiency.
Overlap ratio matters. An overlap ratio of 25% between adjacent patches was found to be optimal. Lower overlap missed contextual information about subtle abnormalities, while higher overlap introduced redundancy without measurable benefit.
Not all patches are useful. When X-ray images are divided into patches, some patches consist entirely or mostly of non-lung regions such as background or bone. Training on these irrelevant patches wastes model capacity and can reduce accuracy.
How the t-test works here. A statistical t-test was applied to each patch to compare pixel intensities between the lung region (group 1) and the non-lung region (group 2). If the difference in mean pixel intensities was statistically significant at a threshold of p less than 0.003, the patch was kept. Otherwise, it was discarded as irrelevant.
Biological rationale. Lung cancer manifests as localized lesions that differ in intensity from surrounding healthy tissue. The t-test identifies patches where such intensity differences exist, which aligns with how radiologists focus their attention on the lung field when diagnosing disease.
Advantage over simpler methods. Compared to adaptive thresholding, the t-test approach proved superior across all performance metrics, achieving 83.19% accuracy versus 79.21% for thresholding. The t-test more effectively isolates discriminative regions relevant to cancer classification.
Five-block design. The custom CNN consists of five sequential blocks, each containing three layers: a 2D convolution layer (Conv2D), a batch normalization (BN) layer, and a dropout layer. After these five blocks, the network has a pooling layer, a flatten layer, a fusion layer, a fully connected layer, and a final softmax classifier.
Feature extraction layers. The Conv2D layer applies 2x2 filters to 256x256 patches, using the ReLU activation function to detect features. Batch normalization standardizes inputs between layers, stabilizing training and providing mild regularization. Dropout layers with a rate of 0.5 prevent overfitting by randomly deactivating neurons during training.
Final classification. The pooling layer reduces feature map size to lower memory requirements and computation time. The flatten layer converts 2D feature maps into a 1D array, which the fully connected layer then maps to class probabilities. The softmax layer produces the final cancer-versus-normal prediction.
A 16-layer network. The proposed CNN has 16 layers total, compared to AlexNet's 8, GoogLeNet's 22, and ResNet-34's 34 layers. This intermediate depth was designed to balance representational power with computational efficiency and resistance to overfitting on small datasets.
Training procedure. All models were evaluated using 10-fold cross-validation on the JSRT dataset. The dataset was split into 10 subsets; each subset was used once as the validation set while the other nine were used for training. This method gives a robust estimate of how the model generalizes to unseen data.
Epoch sensitivity. The proposed CNN achieved its lowest accuracy of 81.9% at 50 training epochs. Performance improved steadily with more epochs, reaching 83.19% at 100 epochs. Performance gains diminished beyond 90 to 100 epochs, indicating a point of diminishing returns.
Outperforming established architectures. When all models were trained using t-test-selected patches, the proposed CNN achieved 83.19% accuracy, outperforming ResNet-34 at 76.5%, GoogLeNet at 71.82%, and AlexNet at 67.62%. The t-test patch selection also reduced variability for all benchmark models compared to full-image training.
Impact of the t-test filter. The difference between using and not using the t-test was dramatic. Without the t-test, the CNN's accuracy dropped to 69.9%, sensitivity fell to 63.46%, and F1-score was only 63.17%. With the t-test applied, accuracy rose to 83.19%, demonstrating the critical importance of the patch filtering step.
Full images slightly outperform patches in raw accuracy. When trained on full CXR images, the proposed CNN reached 84.44% accuracy, marginally higher than the 83.19% achieved with patch-based training. However, this difference was within the standard deviation and not statistically meaningful.
Why patches are preferred in practice. Full-image training requires significantly more GPU memory and longer training times. Patch-based training reduces these resource demands and enables the model to run on standard hardware, which is essential for deployment in low-resource settings.
Generalization advantage of patches. Training on patches increases the effective number of training samples without artificial augmentation such as rotation or flipping. This is particularly valuable when dataset size is small, reducing the risk of overfitting and improving how well the model generalizes to new patients.
Clinical focus on local abnormalities. Whole X-ray images are dominated by healthy tissue, which can dilute the signal from small cancer lesions. Patch-based training allows the model to concentrate on localized abnormalities, which are the key early indicators of lung cancer that radiologists look for.
A practical alternative to CT screening. The proposed CAD system achieves 83.19% accuracy using only chest X-ray images, which cost far less and are more widely available than CT scanners. This makes the system a viable preliminary screening tool for populations in rural or underserved areas.
Outperforming prior JSRT-based methods. Compared to previous studies on the same dataset, the proposed model surpassed all benchmarks: Li et al. achieved 73.92%, Rajagopalan and Babu achieved 72.96%, Ausawalaithong et al. achieved 74.43%, and Ucar and Ucar reached 82.43%. The proposed model at 83.19% represents the new best result on this dataset.
Future directions. The researchers plan to incorporate additional data augmentation techniques, including color space adjustments and kernel filters. Integrating clinical features such as smoking history and family history may further improve diagnostic accuracy. Larger datasets such as NIH Chest X-ray and MIMIC-CXR will be explored to validate performance across diverse populations.