Fusion of Shallow and Deep Features for Classifying Skin Lesions

Abstract

A skin lesion is an unusual change of skin tissues. While this can be caused by harmless skin diseases, there is also the chance of the lesion being cancerous. Skin cancer is one of the most common and deadly cancers in the world, which is caused by exposure to the ultraviolet radiation emitted by the sun. Due to the difficulty in visually differentiating between harmless and cancerous skin lesions, people are less likely to get medical attention straight away. Early diagnosis is crucial to ensure an effective treatment. Clinical and dermoscopy based diagnosis of cancerous skin lesions is costly, painful and sometimes inaccurate. Various researches report performing the classification of skin lesions using image processing techniques. Previous works in this domain are plenty, which reported fairly good results, where image processing and the use of both machine learning and deep learning models are seen. In this research, we propose a novel method which focused on important feature extraction, and fusing multiple features to improve the classification of malignant skin cells using traditional machine learning models, despite having imbalanced data distribution. The ISIC 2018 challenge dataset HAM10000 was used in our work. After preprocessing, we extracted shallow and deep features from the images. Shallow features consisted of position-wise color features and Scale Invariant Feature Transform (SIFT) features. Deep features were extracted by a transfer learning model MobileNetV3, which is pre-trained on Imagenet. These features were combined to form a more representative feature for the data. We parameter tuned five machine learning classifiers to do a binary classification on the processed data. The best accuracy, 81%, was obtained by using Support Vector Machine with an f1-score of 68%. Second best results were achieved by Random Forest Classifier, with an accuracy and F1-score of 80% and 67% respectively.