The Clinical Challenge Lung cancer is the leading cause of cancer-related deaths worldwide. Early and accurate classification of lung cancer type from histopathology images is critical for treatment planning, yet it remains labor-intensive and subject to pathologist variability.
Limitations of Existing Deep Learning Models Traditional convolutional neural networks (CNNs) like ResNet50, DenseNet, and VGG16 achieve high accuracy but suffer from high computational complexity, slow inference times, and tendency to overfit on smaller medical datasets. These limitations make them impractical for real-time clinical deployment in resource-limited settings.
The SCNN Solution This study proposes a Sequential Convolutional Neural Network (SCNN) - a streamlined architecture with three convolutional layers, three max-pooling layers, and dense classification layers. By maintaining a strictly linear flow (no skip connections or parallel branches), SCNN reduces computational complexity while maintaining accuracy.
Three-Class Classification SCNN is trained to distinguish between three categories from histopathological tissue images: lung adenocarcinoma, squamous cell carcinoma, and benign lung tissue. These are the three clinically distinct classes that drive different treatment pathways in non-small cell lung cancer.
Three Data Sources The dataset combined 10,510 histopathological images from three sources: 510 images from Inmol Cancer Hospital Lahore (Pakistan), 3000 images from the National Institute of Health (Pakistan), and 7000 images from Kaggle. This multi-source design improves generalizability over single-center studies.
Balanced Class Distribution Each source contributed equal numbers of adenocarcinoma, squamous cell carcinoma, and benign tissue images, ensuring the model is not biased toward any particular lung cancer subtype.
Data Split Images were divided 70% for training, 15% for validation (holdout), and 15% for testing. The holdout validation strategy ensures hyperparameter tuning does not contaminate final test results.
Preprocessing Pipeline Images were resized from 768x768 to 75x75 pixels to reduce computational load. Pixel values were normalized to the 0-1 range. Gaussian noise layers were added during training to improve model robustness to image quality variation - an important consideration for histopathology images from different scanners and staining protocols.
Three-Layer Convolutional Stack SCNN uses three convolutional layers with filter counts increasing from 32 to 64 to 128 - a pyramid structure that captures increasingly complex features. Each convolutional layer is followed by a max-pooling layer that halves spatial dimensions, reducing computation while preserving key structural information.
Inception V3 Transfer Learning Rather than training from scratch, the model leverages Inception V3 pre-trained weights as the feature extractor. Transfer learning from a model trained on millions of natural images provides a strong starting point and dramatically reduces the amount of labeled medical data required.
Comparison to Complex Architectures Unlike ResNet or DenseNet which use skip connections and dense connections respectively, SCNN maintains a strictly sequential flow. This trades some representational capacity for dramatically lower computational complexity - approximately O(n2 * k2 * d) versus O(n3 * d2) for deeper models.
Ablation Study Confirms Each Component When preprocessing was removed, accuracy dropped from 97% to 85%. Removing feature extraction dropped accuracy to 75%. These ablation results confirm that each architectural component makes a measurable contribution to the final performance.
Overall Performance Across all datasets, SCNN achieved an average accuracy of 95.34%, precision of 95.66%, recall of 95.33%, and an F1-score above 0.95. Training completed within 1000 seconds over 60 epochs on standard consumer hardware (Intel Core i5 with HD Graphics 4600).
Inmol Hospital Dataset On the smallest dataset (510 images), SCNN achieved 0.94 accuracy, precision, recall, and F1-score. Prediction time was 206 milliseconds per step - fast enough for integration into clinical decision-support systems.
Superiority Over Baselines SCNN outperformed standard CNN, R-CNN, and Custom Inception Classifier across all metrics. The improvements were particularly pronounced in F1-score, indicating SCNN's balanced performance across both precision and recall - avoiding the common pitfall of trading one for the other.
Speed Advantage SCNN's reduced computational complexity translates directly to faster inference times. This is critical for clinical adoption, where pathologists need rapid second opinions rather than waiting for extended processing.
Image Downsampling Trade-off Resizing images from 768x768 to 75x75 pixels significantly reduces computational cost but also discards fine-grained morphological details that expert pathologists use for diagnosis. Higher-resolution processing with optimized hardware could recover this information.
Single Cancer Type Focus SCNN is trained exclusively for three lung histology classes. Extending the model to include rare lung cancer subtypes (large cell carcinoma, carcinoid tumors) and to work with other cancer types would increase clinical utility.
Prospective Clinical Integration The model was validated on retrospective datasets. Prospective studies where SCNN assists pathologists in real diagnostic workflows - and where downstream patient outcomes are tracked - are needed to demonstrate genuine clinical benefit.
Planned Next Steps Future work focuses on developing a user-friendly interface for digital pathology integration, running prospective trials at multiple institutions, and exploring further computational optimization to improve performance on more challenging multi-class datasets.