Melanoma Detection using Convolutional Neural Network with Transfer Learning on Dermoscopic and Macroscopic Images

Millenia, Jessica and Naufal, Mohammad Farid and Siswantoro, Joko (2022) Melanoma Detection using Convolutional Neural Network with Transfer Learning on Dermoscopic and Macroscopic Images. Journal of Information Systems Engineering and Business Intelligence, 8 (2). pp. 149-160. ISSN 2443-2555 (online) 2598-6333 (print)

PDF Mohammad Farid Naufal_Melanoma Detection.pdf Download (2MB)

Abstract

Background: Melanoma is a skin cancer that develops when the melanocytes that produce the skin colour pigment start to grow out of control and form cancer. Detecting melanoma early is very important because it dramatically affects patients’ prognosis. A skin examination of a mole indicated as melanoma can be done using dermoscopic or macroscopic images. However, manual screening takes a long time, so automatic melanoma detection is needed. Previous studies still have weaknesses because they yielded low precision or recall. The distribution of melanoma and moles datasets is imbalanced, with the number of melanomas lower than moles. In addition, previous studies have not compared several convolutional neural network (CNN) transfer learning architectures on dermoscopic and macroscopic images. Objective: This study aims to detect melanoma using CNN with transfer learning from dermoscopic and macroscopic melanoma images. CNN with transfer learning is a popular method for classifying digital images with high performance on an imbalanced dataset. Methods: This study compares four CNN with transfer learning architectures, namely MobileNet, Xception, VGG16, and ResNet50, in examining dermoscopic and macroscopic images. This research also uses black-hat filtering and inpainting at the pre-processing stage to remove hair from the skin image. Results: MobileNet is the best model for classifying melanomas or moles in this experiment which has an 83.86% F1 score and 11 seconds of training time per epoch. Conclusion: MobileNet has high average F1 scores of 84.42%, which can detect melanoma accurately even though the number of melanoma datasets is less than moles. It can be concluded that MobileNet is a suitable model for classifying melanomas and moles. In the future, the oversampling method can be implemented to balance the datasets to improve the model's performance.

Item Type:	Article
Uncontrolled Keywords:	CNN, Dermoscopic, Macroscopic, Melanoma, Moles, Transfer Learning
Subjects:	Q Science > QA Mathematics > QA76 Computer software
Divisions:	Faculty of Engineering > Department of Informatic
Depositing User:	MOHAMMAD FARID NAUFAL
Date Deposited:	31 Oct 2022 03:37
Last Modified:	11 Oct 2024 08:32
URI:	http://repository.ubaya.ac.id/id/eprint/42809

Actions (login required)

View Item