Recently, several potentially useful computer-aided diagnosis (CAD) systems have become feasible and are now used widely to help physicians in automated cancer detection and grading from dermoscopy images. In this paper, we present a real-time CAD framework using a deep neural network (DNN) for fine-grained classification and grading in dermoscopic skin cancer images. In this paper, we present a real-time CAD framework using a deep neural network (DNN) for fine-grained classification and grading in dermoscopic skin cancer images. The input skin image is first preprocessed for removing the noise and enhancing the image quality. An adaptive segmentation scheme based on the well-established Otsu thresholding method is performed to accurately extract suspected skin lesion regions from the enhanced input image. Then, a reduced set of visual features is extracted based on both color and typical geometric properties of skin lesions. Finally, the selected lesion features are fed as inputs into a rapid DNN classifier for classifying each lesion in a given dermoscopic image as a benign or melanoma lesion. On the publicly available PH2 dermoscopy imaging dataset, the proposed method is successfully tested and validated, achieving 97.5%, 96.67% and 100.0% for average diagnostic accuracy, sensitivity and specificity, respectively. These results compare quite favorably with those obtained from more sophisticated state-of-the-art approaches.
Digital Object Identifier (DOI)
Bakheet, Samy and El-Nagar, Aml
"A Deep Neural Approach for Real-Time Malignant Melanoma Detection,"
Applied Mathematics & Information Sciences: Vol. 15
, Article 11.
Available at: https://dc.naturalspublishing.com/amis/vol15/iss1/11