Towards Melanoma Detection Using Radar and Image Data

Abstract

Melanoma is a skin cancer type and has the most dangerous consequences. Melanoma spreads to other organs very fast if not detected on time. Several non-invasive techniques exist which are applied in melanoma detection. An example is dermoscopy, which is an optical method and has the advantage of being less costly and easy to use. However, professional expertise is required to detect cancer in the early stage. Artificial Intelligence (AI) has been utilized in skin cancer detection by developing algorithms that can analyse images of skin lesions and identify the characteristics associated with various types of skin cancer, including melanoma. Nevertheless, information about the depth of the melanoma is not provided by the popular technique of using 2D images in training neural networks. The missing depth information is crucial to detecting melanoma and reaching decisions to execute biopsy when necessary. Radar sensors have shown the potential to provide this depth information due to its penetrating capability, allowing them to be applied in the detection of melanoma. The application of AI techniques using 2D images to detect melanoma, and the use of radar, has been investigated independently in recent literature. However, the combined technique still remains to be investigated. We propose integrating radar and image data to improve melanoma classification in this work. Based on the unavailability of radar data, the proposed technique is applied to the skin with nevi and birthmarks, clear skin, and body parts like inner palms, lower arms, and upper arms. The data from both sources are fused by applying an early fusion technique and later utilised for AI classification. Despite the small sample size, the fusion positively impacted classification compared to using only image data. The AI classification was performed on the first two cases, where the overall accuracy increased by 36% for both. Radar signals were also tested on wet and dry skin and have shown distinguishing results.