TransSLC: Skin Lesion Classification in Dermatoscopic Images Using Transformers

• Published in: Lecture notes in computer science Vol. 13413; pp. 651 - 660
• Main Authors: Sarker, Md Mostafa Kamal, Moreno-García, Carlos Francisco, Ren, Jinchang, Elyan, Eyad
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2022; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Computer aided diagnosis; Convolutional neural networks; Deep learning; Skin lesion classification; Transformers
• ISBN: 9783031120527; 3031120523
• ISSN: 0302-9743; 1611-3349; 1611-3349
• DOI: 10.1007/978-3-031-12053-4_48

Early diagnosis and treatment of skin cancer can reduce patients’ fatality rates significantly. In the area of computer-aided diagnosis (CAD), the Convolutional Neural Network (CNN) has been widely used for image classification, segmentation, and recognition. However, the accurate classification of skin lesions using CNN-based models is still challenging, given the inconsistent shape of lesion areas (leading to intra-class variance) and inter-class similarities. In addition, CNN-based models with massive downsampling operations often result in loss of local feature attributes from the dermatoscopic images. Recently, transformer-based models have been able to tackle this problem by exploiting both local and global characteristics, employing self-attention processes, and learning expressive long-range representations. Motivated by the superior performance of these methods, in this paper we present a transformer-based model for skin lesion classification. We apply a transformers-based model using bidirectional encoder representation from the dermatoscopic image to perform the classification task. Extensive experiments were carried out using the public dataset HAM10000, and promising results of $$90.22\%$$ , $$99.54\%$$ , $$94.05\%$$ , and $$96.28\%$$ in accuracy, precision, recall, and F1 score respectively, were achieved. This opens new research directions towards further exploration of transformers-based methods to solve some of the key challenging problems in medical image classification, namely generalisation to samples from a different distribution.