Abstract

Background: The vast majority of the world's population suffers from some form of cancer, which remains one of the most serious global health problems. In 2025 alone, estimating that there will be 20 million new instances of cancer globally, it is predicted that roughly 10.3 million people will die from cancer. Currently used methods to detect cancerous tissues, such as manually analysing mammograms, CT scans and histopathological slides, are limited by factors including variability among different technicians who read the images (known as inter-observer variability), technician fatigue during long hours of training, and delays in diagnosing a patient, which can take between two to four weeks. Integrating AI into oncology has the potential to change the accuracy, scalability and accessibility of the detection process for many types of malignancies (cancers).
Objectives: As the primary objective of this research project, we will develop, train and evaluate an automated system based on a Convolutional Neural Network (CNN) with transfer learning to classify medical images as containing either 1) malignant (cancerous) or 2) non-malignant (non-cancerous) tissue for both breast and lung malignancies. Other secondary objectives of this research will be to augment the automated classification process through the incorporation of Explainable AI (XAI) elements to assist with the clinical interpretation of images as well as explore deployment strategies for resource-limited settings.
Methods: The CBIS-DDSM mammography dataset was used for detecting breast cancer, while the LUNA16 CT dataset was used for detecting lung cancer, by fine-tuning a ResNet-50 model pre-trained on ImageNet. Images were preprocessed through augmentation (rotation, flipping) and resizing (224x224 RGB). The Adam optimizer with binary cross-entropy loss was used via 5-fold CV to train each model over 50 epochs. Grad-CAM was used to enable interpretability of the model. Assessment included using Accuracy, Sensitivity, Specificity, Precision F1-Score and AUC-ROC.
Results: For the CBIS-DDSM dataset, the trained model achieved 94.5% AUC (0.96) accuracy on breast cancer detection and 92.8% AUC (0.94) accuracy on lung cancer detection, outperforming the baseline VGG16 model by an average of 5%-7% across all metrics. Grad-CAM heatmaps were able to successfully show where the tumor was located in relation to clinical significance, and the breast cancer related sensitivity reached 93.2% (reducing critical false negatives).
Conclusion: This CNN-based approach provides exceptional performance in the early detection of multiple types of cancers through imaging. The application of both transfer learning and XAI enables bridging the gap between high precision outputs from DL and clinician trust. As such, the results demonstrate the transformative ability of AI in diagnostics for cancer and should be deployed within healthcare environments that have limited or no specialists.