Detection of Skin Cancer: A Deep Learning Approach

Despite recent advances in medical imaging and machine learning, the detection of skin cancer remains an enormous challenge. Traditional techniques like visual inspection and physical examination by dermatologists are prone to subjectivity and heavily dependent on the clinician’s experience, resulting in variable diagnostic accuracy. This manifests the pressing need for more accurate, efficient and available diagnostic techniques. Deep learning techniques, particularly transfer learning, have demonstrated promise in addressing problems in multiple disciplines, especially in medical imaging domain. This research seeks to address the issue by developing a transfer learning model for skin cancer detection from images using pre-trained deep learning architectures such as: DenseNet201, ResNet50, and VGG16 with task-specific additional layers. The developed skin cancer detection model attained an aggregate validation accuracy of 91% with F1-scores ranging from 0.94 to 1.00 for some attributes which is highly accurate, efficient, and potentially more accessible than current methods.