Effects of objects and image quality on melanoma classification using Deep Neural Networks

Abstract

Abstract Background: Melanoma is a type of skin cancer with a higher mortality compared to other types of skin cancers. Early and accurate diagnosis of melanoma has critical importance on its prognosis. Recently, deep neural network based models dominated the CAD systems for classification of the potential melanoma lesions. In clinical settings, capturing impeccable skin images is not always possible. In some cases, an external object such as a ruler is required for determination of lesion size. Sometimes, the skin images can be blurry, noisy or have low contrast. The aim of this work is to investigate the effects of external objects (ruler, hair) and image quality (blur, noise, contrast) on the classification of melanoma using commonly used Convolutional Neural Network(CNN) models.Results: Performance is analyzed using accuracy, sensitivity, specificity and precision metrics over 6 different test sets. Hair set has 89.22%, ruler set has 86% and none set has 88.81% as the best accuracy with DenseNet121 architecture. Also, DenseNet has the best average accuracy with comparing the other three models in other datasets, which are noise and blur. We find that ResNet is better for contrast dataset. We can infer that DenseNet can be used for melanoma classification with image distortions and degradations. Conclusion: In this study, we investigate the effect of ruler/hair and image blur, noise and contrast on the melanoma detection performance of four commonly used CNN models: ResNet50, DenseNet121, VGG16 and AlexNet. Melanoma images can be better recognized under contrast changes unlike the benign images, we recommend ResNet model whenever there is contrast issue. Noise significantly degrades the performance on melanoma images and the recognition rates decrease with compared to benign lesions in noisy set. DenseNet121 also works well in this set. Both classes are sensitive to blur changes and best accuracy is obtained with DenseNet model. The images contain ruler has decreased the classification accuracy and ResNet has better performance if there is ruler in an image. Hairy images have the best success rate in our system since it has the maximum number of images in total dataset. DenseNet performs better for both hairy and high quality images.