DePicT Melanoma Deep-CLASS: a deep convolutional neural networks approach to classify skin lesion images

Abstract

Background: Melanoma results in the vast majority of skin cancer deaths during the last decades, even though this disease accounts for only one percent of all skin cancers' instances. The survival rates of melanoma from early to terminal stages is more than fifty percent. Therefore, having the right information at the right time by early detection with monitoring skin lesions to find potential problems is essential to surviving this type of cancer.

Results: An approach to classify skin lesions using deep learning for early detection of melanoma in a case-based reasoning (CBR) system is proposed. This approach has been employed for retrieving new input images from the case base of the proposed system DePicT Melanoma Deep-CLASS to support users with more accurate recommendations relevant to their requested problem (e.g., image of affected area). The efficiency of our system has been verified by utilizing the ISIC Archive dataset in analysis of skin lesion classification as a benign and malignant melanoma. The kernel of DePicT Melanoma Deep-CLASS is built upon a convolutional neural network (CNN) composed of sixteen layers (excluding input and ouput layers), which can be recursively trained and learned. Our approach depicts an improved performance and accuracy in testing on the ISIC Archive dataset.

Conclusions: Our methodology derived from a deep CNN, generates case representations for our case base to use in the retrieval process. Integration of this approach to DePicT Melanoma CLASS, significantly improving the efficiency of its image classification and the quality of the recommendation part of the system. The proposed method has been tested and validated on 1796 dermoscopy images. Analyzed results indicate that it is efficient on malignancy detection.

Keywords: Case-based reasoning; Classification; Deep learning; Early detection; Information retrieval; Melanoma; Skin cancer.

Fig. 1

Case representation. Case structure of DePicT Melanoma CLASS [10]

Fig. 2

Overview of the proposed CBR system. DePicT Melanoma CLASS enriched with CNN method, adopted from [11]

Fig. 3

MAX pooling. forward propagation [36]

Fig. 4

Different Activation Functions. a ReLU and Leaky ReLU [37], b Sigmoid Activation Function [37], c Step Activation Function [38]

Fig. 5

Overview of the CNN. Layer view and a MNIST digit classification example of a Caffe network - Sixteen layers excluding max-pooling [11]

Fig. 6

Complete layer view and a MNIST digit classification example of our Caffe network. Sixteen layers excluding max-pooling; convolution in red, max-pooling in orange and fully connected layers in violet

Fig. 7

Sample images. Benign and malignant melanoma images from IS IC Archive dataset [22]. a B b MM

Fig. 8

Evaluation results of DePicT Melanoma Deep-CLASS (CNN) in comparison with DePicT Melanoma CLASS. ROC and Precision-recall curves in the first and third round of training-testing: a k-NN, b SVM, and c CNN

Fig. 9

Evaluation results of DePicT Melanoma Deep-CLASS (CNN) in comparison with DePicT Melan oma CLASS. ROC and Precision-recall curves in the second and fourth round of trainingtesting: a k-NN, b SVM, and c CNN