Explainable 3D Convolutional Neural Networks for Enhancing Early Lung Cancer Detection

Abstract

Lung cancer is a major cause of cancer-related deaths, often due to late diagnosis. Early detection is vital for better outcomes and lower mortality rates. Traditional methods like chest X-rays are not sensitive enough to spot small, early-stage nodules, highlighting the need for advanced imaging and classification techniques to identify malignant nodules early on. This study investigates the application of interpretable deep learning models for classifying lung nodules in CT scans as benign or malignant, aiding in the early detection of lung cancer. We leverage 3D Convolutional Neural Networks (3D CNNs) trained on the LUNA22 ISMI dataset, comprising 1,176 lung nodules with a collection of lung nodule annotations from anonymized CT scans. In the preprocessing stage, we standardized pixel intensity values and applied augmentation techniques such as rotation, scaling, and flipping to enhance the diversity of the training data. These steps were implemented using Python packages, including SciPy for augmentation and NumPy for array operations. We compared four 3D CNN models with different architectures, including a baseline model, a 3D AlexNetbased model, a proposed 3D CNN model, and a 3D CNN model integrated with the Convolutional Block Attention Module (CBAM). The CBAM module enhances feature extraction by applying attention mechanisms to the most informative features. Our proposed 3D CNN with CBAM achieved the highest performance, with an accuracy of 94.06%, AUC of 98.84%, and F1-Score of 95.56%. To ensure model transparency and facilitate clinical adoption, we employed 3D Grad-CAM to generate visual explanations of the model's predictions. This technique provides insights into the regions of the lung nodule that most influence the classification decision. Despite these promising results, the study has several limitations. The dataset used, although robust, is limited in size and diversity, potentially impacting generalizability to broader populations. Additionally, our models were trained and evaluated under controlled conditions, which may not fully replicate real-world clinical environments. Future work should focus on validating these models in larger, more diverse datasets and deploying them in prospective clinical studies to assess their practical impact. Integration with existing diagnostic workflows and evaluating cost-effectiveness are also critical steps toward clinical adoption.