Global burden of lung cancer. Lung cancer is the leading cause of cancer-related deaths worldwide, responsible for approximately 1.8 million deaths in 2020 alone. With 2.48 million new cases reported in 2022, it is one of the most common cancers globally, driven by tobacco use, aging populations, and lifestyle changes in developing countries.
Non-small cell lung cancer dominates. Non-small cell lung cancer (NSCLC) accounts for approximately 85% of all lung cancer cases and includes three primary subtypes: adenocarcinoma (AC), squamous cell carcinoma (SCC), and large cell carcinoma (LCC). Each subtype has distinct clinical and molecular characteristics that affect treatment decisions.
Limitations of current diagnostic methods. Traditional methods such as biopsy and imaging are often costly and depend heavily on the expertise of radiologists. CT scans are widely used but diagnostic accuracy can vary. There is a clear need for reliable, automated tools that provide consistent and accurate lung cancer subtype classification.
AI as a clinical decision support tool. Deep learning-based clinical decision support systems have shown promise in assisting radiologists by providing more reliable and consistent predictions. However, most state-of-the-art models are computationally intensive and require high-end hardware, limiting their use in resource-constrained clinical settings.
A lightweight architecture for real-world use. The researchers developed MiniConvNet, a novel lightweight convolutional neural network (CNN) designed specifically to balance diagnostic accuracy with computational efficiency. It was developed through systematic hyperparameter optimization and iterative refinement to maximize sensitivity and specificity while minimizing computational demands.
Dramatically fewer parameters. MiniConvNet contains only 0.5 million trainable parameters and has a model size of just 6 MB. In comparison, standard models like VGG16 have 117.5 million parameters and a file size of 1320 MB. This compact design enables MiniConvNet to run on low-resource hardware, including embedded systems and edge devices in clinical environments.
Design principles behind the architecture. The model uses convolutional layers as building blocks for hierarchical feature extraction, interleaved with ReLU activation functions, max pooling operations for dimensionality reduction, and batch normalization for training stability. A flattening layer and dense layers connect to a SoftMax output for multi-class classification.
Designed for multiclass NSCLC classification. Unlike many prior studies focused on binary cancer detection or nodule identification, MiniConvNet addresses the clinically important problem of distinguishing between AC, SCC, LCC, and healthy lung tissue from CT images. It was also evaluated on histopathological images to test generalizability across imaging modalities.
Two publicly available datasets. The study used two datasets: a lung CT scan dataset from Kaggle containing 900 images across four classes (338 AC, 260 SCC, 187 LCC, and 115 healthy), and a histopathological images dataset with 15,000 images across three classes (AC, SCC, and healthy lung). Using two datasets allowed cross-modality evaluation of the model.
Clinical staging details of CT images. The CT dataset captures clinically staged tumors: AC lesions are T2-stage located in the left lower lobe, LCC tumors are near the left hilum with lymph node metastasis, and SCC tumors are T1-stage at the lung hilum also with lymph node involvement. Healthy images show normal lung tissue with no abnormalities.
Addressing the small dataset challenge. With only 900 CT images, there is a significant risk of overfitting. To mitigate this, the researchers applied extensive data augmentation during training, including horizontal and vertical flipping, random rotations, brightness adjustments, and zoom transformations, which effectively increased training data variability.
Standardized preprocessing and splits. Both datasets were divided into 80% training, 10% validation, and 10% testing subsets. Images were resized to 224 x 224 pixels and pixel values were rescaled to the range [0,1] using Keras ImageDataGenerator. This standardized preprocessing ensured consistency across both CT and histopathology data.
Seven established architectures as benchmarks. MiniConvNet was benchmarked against ResNet-50, VGG16, VGG19, Inception V3, MobileNetV3Small, EfficientNetV2B0, and ConvNeXtTiny. These models were all fine-tuned on the same lung cancer datasets under identical experimental conditions to ensure a fair comparison.
Architectural diversity of baselines. The baseline models range from deep residual networks (ResNet-50 with skip connections) to inception-style parallel feature extraction (Inception V3), lightweight depthwise separable convolution models (MobileNetV3), and transformer-inspired designs (ConvNeXtTiny). Together they represent the range of modern CNN approaches used in medical imaging.
Unified training protocol. All models were trained using the Adam optimizer with categorical cross-entropy loss, a batch size of 16, and early stopping with a patience of 10 epochs monitoring validation loss. The number of training steps per epoch was calculated from the training set size divided by the batch size, ensuring complete coverage in each epoch.
Evaluation metrics. Model performance was measured using accuracy, precision, recall, and F1-score, calculated from true positives, false positives, and false negatives. Loss values were also compared. This multi-metric evaluation allowed a thorough comparison of diagnostic performance and computational efficiency across all eight models.
96% accuracy versus 82% for the next best model. On the CT scan dataset, MiniConvNet achieved 96% test accuracy, the highest among all evaluated models. The next best was Inception V3 at 82%, followed by VGG16 at 77%, VGG19 at 69%, ConvNeXtTiny at 58%, ResNet-50 and EfficientNetV2 both at 52%, and MobileNetV3 at 46%.
Consistent performance across all metrics. MiniConvNet achieved precision, recall, and F1-score of 0.96 each on the CT dataset, with the lowest loss at 0.15. By contrast, Inception V3 had a loss of 0.42. Several models like EfficientNetV2 and ConvNeXtTiny showed high precision but very low recall, indicating conservative predictions and poor sensitivity.
Strong performance on histopathology data. MiniConvNet achieved 96.8% test accuracy on the histopathological image dataset, confirming its ability to generalize across imaging modalities. On the training and validation sets it reached 99.1% and 98% accuracy respectively, with precision and recall of 0.98 on both, demonstrating consistent classification performance.
Class-wise performance details. On the CT dataset, MiniConvNet achieved perfect recall and precision of 1.00 for the healthy lung class. AC achieved 0.94 recall and 0.97 precision, and SCC had slightly more misclassifications, which is consistent with the clinical difficulty of distinguishing SCC from other subtypes. On histopathology, LCC and healthy lung classes showed particularly strong performance.
Faster training with lower memory footprint. MiniConvNet required 133,000 ms training time on the CT dataset, less than VGG19 (210,000 ms), EfficientNetV2B0 (250,000 ms), and ConvNeXtTiny (200,000 ms). Its inference time of 5 ms is competitive with faster models like EfficientNetV2 (4 ms) and MobileNetV3 (5 ms), while its model size of 6 MB is the smallest by a large margin.
Dropout improves generalization. To assess regularization, the researchers compared MiniConvNet with and without dropout layers between dense layers. The version with dropout trained for 17 epochs (versus 23 without) and achieved higher validation accuracy of 0.9710 with a lower validation loss of 0.1750, indicating better balance between training and validation performance.
Five-fold cross-validation confirms robustness. A stratified five-fold cross-validation experiment was conducted to rule out overfitting on a single data split. Across all five folds, test accuracies ranged from 94.1% to 97.4%. The mean performance across folds was 96.6% accuracy, 96.6% precision, 96.4% recall, and 96.5% F1-score, confirming stable generalization.
Minimal variation between folds. The small variation in performance across folds (less than 3.3 percentage points difference between the lowest and highest accuracy folds) reflects natural differences in data distribution rather than model instability. This consistency demonstrates that MiniConvNet is not overfitting and performs reliably across different training and test splits.
Gap between training and test accuracy. While MiniConvNet achieved 100% training accuracy, its validation and test accuracy settled at 96%, a drop that suggests some sensitivity to shifts between training and real-world data distributions. This is a common challenge in medical imaging when datasets are small or not fully representative of clinical diversity.
Risk of overfitting on small datasets. The CT dataset of only 900 images is significantly smaller than what is typically preferred for deep learning. Although data augmentation and early stopping were applied, the risk of overfitting remains, particularly if the model were applied to CT images from different scanners, patient populations, or imaging protocols.
Computational efficiency has practical limits. Although MiniConvNet is far more efficient than deeper architectures, real-time clinical deployment may still require careful optimization depending on the hardware infrastructure available. In settings with very limited computing power, even lightweight models may need further compression or quantization.
Future work recommendations. The authors suggest validating MiniConvNet on larger and more diverse datasets, potentially integrating transfer learning techniques to improve generalization. Exploring how the model can be integrated directly into clinical workflows, such as as a real-time decision support layer for radiologists, is also identified as a key next step for translation to practice.
A meaningful step toward deployable lung cancer AI. MiniConvNet represents a practical advancement in AI-assisted lung cancer diagnostics by achieving both high accuracy and computational efficiency. Its 96% accuracy on CT scans and 96.8% on histopathology images, combined with a model size of only 6 MB, makes it a strong candidate for real-world clinical deployment.
Addressing a critical gap in AI diagnostics. Many existing models prioritize accuracy at the expense of computational resources, making them impractical for hospitals with limited infrastructure. MiniConvNet directly addresses this gap by providing a lightweight yet high-performing alternative that does not require expensive GPUs or large memory allocations.
Superiority over well-established architectures. The study demonstrates that a carefully designed lightweight architecture can outperform complex state-of-the-art models. MiniConvNet surpassed ResNet-50, VGG16, VGG19, Inception V3, MobileNetV3, EfficientNetV2, and ConvNeXtTiny on NSCLC classification, achieving the highest accuracy and lowest loss among all tested models.
Broader implications for cancer AI. The success of MiniConvNet across two imaging modalities (CT and histopathology) suggests that the lightweight design principles behind it could be applied to other cancer types and imaging domains. This work contributes to the larger goal of making AI-driven diagnostic tools accessible and clinically viable across diverse healthcare settings.