The Intelligent Systems Research Group (ISysRG) has completed research on the classification of skin lesions. One of the potentially cancerous skin lesions is melanoma (Mayo Foundation for Medical Education and Research (MFMER), 2017). Melanoma is formed in melanocyte cells (Hartman & Lin, 2019). Today, melanoma cancer causes many deaths in the world (American Cancer Society. Atlanta, 2019). Melanoma examination in the early stages is very important so that patient survival can be improved (Balch et al., 2009). One way to detect melanoma cancer is to use digital imaging called dermoscopy. Medical diagnosis using a dermoscopy with a computerized system is needed at this time.

We have implemented the classification of skin lesions in the HAM10000 dataset. This dataset consists of 10015 images of skin lesions with 7 diagnosis categories. This dataset can be accessed publicly at Samples of dermoscopy image data in each diagnostic category can be seen in Figure 1

Figure 1. Dermoscopy Image Samples on HAM10000 Dataset

Deep learning (DL) has been used to process dermoscopy image data in several studies (Esteva et al. 2017) (Romero Lopez et al. 2017) (Yamashita et al. 2018) (Mahmud et al. 2018) (Hosny, Kassem, and Foaud 2019) (Shahin et al., 2019). The DL method applied for the classification of skin lesions is Convolutional Neural Networks (CNN). CNN is generally used to process 2D data, for example, a color images consisting of 2D arrays (LeCun et al., 2015). CNN is composed of convolutional layers, pooling layers, and fully connected layers (Zhao & Kumar, 2017). The ilustration of CNN architecture for skin lesion classification can be seen in Figure 2.

Figure 2. CNN Architecture

Our research proposes an ensemble model approach by combining three deep CNNs architectures namely Inception V3, Inception ResNet V2, and DenseNet 201 which can be seen in Figure 3. Our system can classify 7 classes of skin lesions with excellent results. This research achieved average accuracy, average precision, average sensitivity, average specificity, and average F1 scores are 97.23%, 90.12%, 97.73%, 82.01%, and 85.01% respectively

Figure 3. Proposed CNNs with Ensemble Learning

Tags: No tags

5 Responses

Add a Comment

Your email address will not be published. Required fields are marked *