2023 International Conference on Electrical, Computer and Communication Engineering (ECCE) SkinNet-8: An Efficient CNN Architecture for Classifying Skin Cancer on an Imbalanced Dataset Nur Mohammad Fahad Department of CSE United International University (UIU) Dhaka, Bangladesh nfahad191040@bscse.uiu.ac.bd Sadman Sakib Department of CSE United International University (UIU) Dhaka, Bangladesh ssakib191097@bscse.uiu.ac.bd Mohaimenul Azam Khan Raiaan Department of CSE United International University (UIU) Dhaka, Bangladesh mraiaan191228@bscse.uiu.ac.bd Md. Saddam Hossain Mukta Department of CSE United International University (UIU) Dhaka, Bangladesh saddam@cse.uiu.ac.bd Abstract—Skin cancer is a fatal disease that has become the leading cause of death worldwide in recent years, although it is curable if diagnosed early. Early skin cancer detection significantly improves patients’ chances of survival and reduces mortality. In this research, we conduct experiments on a high imbalance dermoscopic ISIC 2020 dataset. The primary objective of this study is to develop a shallow CNN architecture to complete the classification task effectively, requiring fewer computational resources without compromising accuracy. We have used pre- processing techniques to remove image noise and truncation and augmentation techniques to balance the dataset before fitting it into the model. Multiple performance measurement metrics were utilized to establish the overall performance. Our proposed model yields a remarkable test accuracy of 98.81%. We compare our models’ performance with different transfer learning (TL) models to assess the faster convergence rate. The proposed model demonstrated its robustness by outperforming the other TL models in terms of accuracy within a short processing time. It is reasonable to assume that our proposed system will reliably aid dermatologists in diagnosing skin cancer patients early and increasing survival rates. Index Terms—Skin cancer, deep Learning, CNN, image pre- processing, data balance, shallow I. I NTRODUCTION Skin cancer is one of the most severe threats to global health. In 2020, approximately 1,198,073 new cases of skin cancer and 63,731 deaths had been reported from skin cancer [1]. Melanoma and non-melanoma are the two categories of skin cancer [2]. Among the two types of skin cancer, melanoma skin cancer is more menacing as 75% of deaths from skin cancer are due to malignant melanomas [1]. It is caused by the abnormal growth of melanocyte cells, which tend to replicate and spread through lymph nodes, destroying the surrounding tissue [3]. Lesions appear on the epidermis, the outermost layer of skin, during the early stages of skin cancer development. Late-stage skin cancer is more likely to spread and is more challenging to treat. Therefore, early de- tection and diagnosis of skin cancer are essential to preventing the spread of cancer, lowering the mortality rate of skin cancer, and enhancing global survival rates [4]. Clinicians frequently conduct screenings for skin cancer via visual examination, which is less accurate and less objective than other methods [5]. It is essential to build a dependable automatic system for detecting skin cancer that improves pathologists’ precision and productivity. Dermoscopy is a technology that has been developed to enhance the diagnosis of skin cancer. Dermoscopy is a noninvasive skin imaging procedure that involves capturing a magnified and lit image of the skin region for greater clarity of the spots, which improves the visual effect of a skin lesion by reducing surface reflection [6]. The diagnostic efficacy of dermoscopy relies on the dermatologist’s expertise and training. However, der- matologists typically identified melanoma from dermoscopic images with less than 80% accuracy in clinical settings [7]. An automated diagnosis system can facilitate the advancement of clinical decisions concerning skin cancer detection. With this aim, our study uses dermoscopy to classify skin cancer accurately. In recent years, automated diagnostic tools based on deep learning have garnered significant attention due to their re- markably improved prediction accuracy [8]. Reis et al. [9] proposed a deep convolutional method, InSiNet, for detecting and segmenting skin cancer for the ISIC 2020 dataset [10]. The accuracy of their study was 90.54%, and a comparison of their model to other transfer learning models revealed that the InSiNet architecture yields the best overall results. A recent study by Kaur et al. [11] propose a deep CNN-based automated melanoma classifier that can distinguish between malignant and benign Melanoma with high accuracy. On the ISIC 2020 dataset, their proposed model surpassed transfer learning models ResNet18, Inceptionv3, and AlexNet with an accuracy of 90.42%. In another recent study, Elansary et al. [12] introduce a CNN-based classification model for classifying Melanoma on the ISIC 2020 dataset. The model classified patients’ skin lesions as malignant or benign uti- 979-8-3503-4536-0/23/$31.00 ©2023 IEEE 2023 International Conference on Electrical, Computer and Communication Engineering (ECCE) | 979-8-3503-4536-0/23/$31.00 ©2023 IEEE | DOI: 10.1109/ECCE57851.2023.10101527 Authorized licensed use limited to: Charles Darwin University. Downloaded on April 21,2023 at 05:54:49 UTC from IEEE Xplore. Restrictions apply.