MetaAttention model: a new approach for skin lesion diagnosis using AB features and attention mechanism

• Published in: Biomedical physics & engineering express Vol. 9; no. 4; pp. 45008 - 45029
• Main Authors: To, Hung Duy, Nguyen, Huy Gia, Le, Hang Thi Thuy, Le, Hung Minh, Quan, Tho Thanh
• Format: Journal Article
• Language: English
• Published: England IOP Publishing 01.07.2023
• Subjects: attention mechanism; deep learning; Humans; image processing; Male; Melanoma - diagnosis; Skin - pathology; skin cancer; Skin Diseases; Skin Neoplasms - diagnosis; Skin Neoplasms - pathology; transfer learning; Ultraviolet Rays; image processing; deep learning; skin cancer; transfer learning; attention mechanism
• ISSN: 2057-1976; 2057-1976
• DOI: 10.1088/2057-1976/acd1f0

Ozone depletion has always been a hot crisis around the globe. Its consequence is the increase in ultraviolet radiation at the surface in many regions and countries, which then causes danger to the human immune system, eyes, and especially skin - the part that is directly exposed most to the sunlight. According to the World Health Organization, the number of cases of skin cancer is higher than that of breast, prostate, and lung cancer combined. Therefore, there has been a lot of research to apply deep learning models in solving the problem of skin cancer classification. This paper proposes a novel approach, named MetaAttention, aimed at improving the performance of transfer learning models for skin lesion classification. The method combines image features with patients' metadata features using the attention mechanism, incorporating clinical knowledge related to ABCD signals to better distinguish melanoma cell carcinoma disease, which has long been a significant challenge for researchers. The experimental results indicate that the proposed approach outperforms the state-of-the-art method, EfficientNet-B4, achieving an accuracy of 89.9% with Scale-dot product MetaAttention and 90.63% with Additive MetaAttention. The method has the potential to support dermatologists in diagnosing skin lesions effectively and efficiently. Furthermore, with larger datasets, our method could be further fine-tuned to achieve even better performance on a broader range of labels.