Volume 18, No. 6, 2021

Identification Of Multi-Class Skin Cancer Classification Using Relation Based Embedding Network With Few Shot Learning

Gopikha S and Balamurugan M


In order to have the best possible chance of survival for patients, early identification of skin malignancies such as melanoma is critical. Patient care could be improved through clinical use of Deep Learning-based Decision Support Systems for skin cancer screening. Many medical AI researchers focus on a diagnosis scenario that is primarily useful for autonomous operation. This is the primary emphasis of medical AI research. It is important, however, that practical decision support go beyond merely diagnosing and explain why. For dermatological disease diagnosis, the issue of clinical image classification is examined. Relation embedding networks are used to solve the challenge of few-shot classification, when a classifier has to generalise to classes that it has never seen before, with only a few examples of each novel class to work with few shot learning and classification can benefit from a relation-embedded network. Modules such as embedding metric mapping for feature vector evaluation and location detection for lesion identification during testing have been proposed in this work as two different modules. Using publicly obtainable skin lesion datasets, such as the SD-198 dataset, we evaluate the network's effectiveness and find that it outperforms than existing techniques with only a few annotated instances.

Pages: 5520-5536

Keywords: Skin Cancer; Relation-Embedding Networks; Image Classification; Few Shot Learning; Decision Support System.

Full Text