An Enhanced Technique for Skin Lesion Diagnosis using Dermoscopic Images

Mosa, Aya Mostafa and Afifi, Ahmed and Amin, Khalid (2022) An Enhanced Technique for Skin Lesion Diagnosis using Dermoscopic Images. IJCI. International Journal of Computers and Information, 9 (2). pp. 74-87. ISSN 2735-3257

[thumbnail of IJCI_Volume 9_Issue 2_Pages 74-87.pdf] Text
IJCI_Volume 9_Issue 2_Pages 74-87.pdf - Published Version

Download (588kB)

Abstract

There are many types of skin cancer, the most harmful type among them is melanoma. Skin cancer occurs through the abnormal growth of the body cells, which may be caused by continuous exposure to ultraviolet rays resulting from the sun. Early diagnosis of skin cancer is essential as it can reduce the burden, make the treatment more effective and save the patient life. In this work, therefore, we develop an enhanced ensemble method to improve the classification accuracy of eight types of skin cancer. Transfer learning of three pretrained Convolutional Neural Network (CNN) models, namely resnet18, densnet121, and inception v4, are used as a base for this ensemble. Firstly, we fine-tune each pre-trained model separately on an augmented dataset. Afterward, different ensemble methodologies are applied including average ensemble, ensemble using support vector machine (SVM), and random forest (RF) classifiers. The ensemble method using the SVM, and RF improved the accuracy of the average ensemble method by combining the prediction of pre-trained CNN models as input to SVM and RF classifiers. The pre-trained models were fine-tuned and evaluated using 17731 and 3800 images from different types of skin cancer. The individual pretrained models of Resnet18, Densnet121, and InceptionV4 achieved an accuracy of 79.5%, 81.2%, and 82.6% respectively. The proposed ensemble method using SVM, and RF classifiers gives the best accuracy result with 85% for the SVM classifier and 86.2% for the Random Forest classifier. Results show that the proposed ensemble method using SVM, and RF classifiers outperform the individual pretrained models.

Item Type: Article
Subjects: STM Digital > Computer Science
Depositing User: Unnamed user with email support@stmdigital.org
Date Deposited: 07 May 2024 05:29
Last Modified: 07 May 2024 05:29
URI: http://research.asianarticleeprint.com/id/eprint/1362

Actions (login required)

View Item
View Item