Deep learning driven image-based cancer diagnosis

Lung cancer remains the leading cause of cancer-related mortality worldwide, primarily due to late-stage detection, which limits treatment availability. Early detection with low-dose CT screening increases the chance of survival, yet interpretation of CT scans for subtle malignant nodules is both difficult and time-consuming for radiologists. We present a new CAD system that utilizes deep learning techniques in the clinical imaging workflow for the detection of early-stage lung cancer. Our method adopts a hybrid CNN-Transformer model, which is designed to simultaneously learn local nodule properties and global contextual patterns based on attention mechanisms. The system is intended to be clinically feasible and has been constructed with the capability to interface seamlessly with hospital PACS and DICOM for image retrieval and storage. We address the most common forms of bias in imaging AI by applying data normalization, class-balanced data augmentation, and a bias-mitigating training scheme to generalize across different imaging devices and patient subpopulations. Experiments on public datasets (LIDC-IDRI for lung nodules and DeepLesion for diverse lesions) and a simulated multi-institution dataset show that the model achieves high accuracy and generalization. Notably, the hybrid CNN-Transformer demonstrates better performance across all baseline classifiers (including conventional radiomics SVM, CNN, or Vision Transformer models), with an overall AUC of 0.97 (95% confidence interval: 0.94–0.99), a sensitivity of 95.1%, and a specificity of 96.5% on the independent test set. Statistical tests (DeLong’s test for AUC) show that our method is significantly better than previous approaches (p<0.01). We provide complete technical material, including model architecture, training pseudocode, and all evaluation scores with confidence intervals and p-values. Furthermore, we demonstrate qualitative results with Grad-CAM visualizations of image regions most pertinent to predicting classification labels, thereby providing interpretability for clinicians. We also address responsible AI practices such as bias audits at the subgroup level (e.g., age, sex, scanner type), documentation via model cards for transparency, adversarial robustness testing, and compliance with privacy regulations (de-identification and Data Protection Impact Assessment). Our results indicate that the bias-aware hybrid CNN-Transformer is a powerful yet reliable solution for image-based lung cancer diagnosis and is a promising enabler of earlier diagnosis and better patient outcomes, while complying with ethical and security considerations of medical AI.