Chest Cancer Classification from Chest CT-Scan Images using Deep Learning

Chest cancer is fatal, encompassing various tumors. This includes lung cancer as the most common, originating from tissues of the lungs. All chest cancer types present unique challenges, leaving individuals and healthcare professionals searching for solutions. The conventional fight against chest cancer includes surgery to remove the cancerous tissue, radiation therapy to target remaining cancer cells, and chemotherapy to attack cancer cells throughout the body. While these methods have seen advancements, innovative strategies are still required to aid doctors in classifying the different chest cancer types. Deep learning is currently showing potential, where researchers are developing sophisticated algorithms capable of analyzing medical images, identifying subtle patterns, and distinguishing between various cancer types. This research investigates the application of DenseNet121, a deep learning architecture with a combination of 2 data augmentation techniques namely, albumentations and mixup for classifying chest cancer types based on chest CT-scan images. Four image types were classified including adenocarcinoma, large cell carcinoma, squamous cell carcinoma and normal. The performance of the trained model for this research was based on accuracy, precision, recall and F1-score parameters. It was noted that DenseNet121 achieved an accuracy of 82.14% while combining albumentations and mixup, 78.64% with only mixup, 77.78% with only albumentations and 58.69% without data augmentation, indicating that the proposed model performed better. Combining albumentations and mixup took advantage of the strengths of both augmentation techniques to increase the size and the diversity of training datasets, thereby favoring regularization to reduce the overfitting challenge faced by deep learning models. However, further research is still required to evaluate the findings of this research on larger and other kinds of datasets and most importantly, compare them to the real-world scenarios that might introduce variations with a wider variety of cases.