International Journal of Computer Applications (0975 – 8887) Volume 186 – No.18, April 2024 17 Skin Lesion Prediction from Dermoscopic Images using Deep Learning Nazma Hossen Nishat Dept. of Computer Science and Engineering Port City International University Chittagong, Bangladesh Pranta Paul Dept. of Computer Science and Engineering Port City International University Chittagong, Bangladesh Farzina Akther Dept. of Computer Science and Engineering Port City International University Chittagong, Bangladesh Tahmina Akter Dept. of Computer Science and Engineering Port City International University Chittagong, Bangladesh Muhammad Anwarul Azim Dept. of Computer Science and Engineering University of Chittagong Chittagong-4331, Bangladesh ABSTRACT Skin lesions, which comprise a wide range of irregularities in skin appearance, might serve as precursors of skin cancer due to the complex interaction of hereditary variables and longterm UV ex- posure. Significant advances in dermatology have been made with the use of deep learning models, notably convolutional neural net- works (CNNs). These models excel in analyzing dermatoscopic pictures, allowing for early and accurate identification of a vari- ety of skin problems. In this work, a complete evaluation of deep learning models for predicting skin lesions is conducted, with an emphasis on accuracy. Notable performers include DenseNet169 and ResNet101, both of which achieve an outstanding 91% accu- racy. Furthermore, a hybrid model obtains an accuracy of 89%, in- dicating its capacity to recognize complicated visual patterns. The study investigates model fusion strategies to capitalize on possible synergy in prediction skills, ultimately improving automated der- matological diagnosis systems. Notable models are DenseNet121, ResNet-50V2, and InceptionResNetV2, which contribute consider- ably with accuracies of 91%, 89%, and 85%, respectively, while MobileNetV2 and VGG-16 provide accuracies of 82% and 80%. These advances, taken together, enable the development of strong and accurate diagnostic technologies capable of efficiently expedit- ing skin health interventions. Keywords CNN,Transfer Learning, Data Balancing, Augmentation, Hybird Model 1. INTRODUCTION The area of dermatology has seen dramatic advances in skin le- sion prediction, owing to the use of deep learning (DL) approaches, notably those based on dermoscopic pictures. Skin lesions, which vary in appearance, frequently cause concerns owing to their po- tential link with skin cancer. Recognizing the significance of early detection, academics have embraced deep learning (DL), specifi- cally convolutional neural networks (CNNs), to improve diagnos- tic capabilities. These models excel in analyzing dermoscopic pic- tures, allowing for accurate detection and categorization of numer- ous skin conditions. The predictive value of these models lies from their capacity to detect minor patterns and features associated with certain lesions, allowing for early intervention. Automated predic- tion of skin lesions utilizing deep learning models, such as CNN, represents a paradigm leap in dermatological diagnostics by dra- matically lowering analysis time while also giving a reliable and accurate way of identifying possible issues. This invention not only complements existing diagnostic approaches, but also has the po- tential to improve patient outcomes through prompt treatments. As deep learning advances, a thorough evaluation of various mod- els, such as DenseNet169, ResNet101, and hybrid combinations, on distinct dermoscopic datasets becomes increasingly important. Such assessments, which take into account performance measures as accuracy, not only highlight the strengths of certain models but also direct the investigation of model fusion approaches for possi- ble synergies. In summary, the combination of deep learning with dermoscopic imaging ushers in a new era in dermatology, promis- ing improved predictive capacities and considerably contributing to the early identification and management of skin diseases. 2. LITERATURE REVIEW According to the research of Neeshma, A. et al., [1] multi-class skin lesions may be classified into seven groups using a deep learning technique that uses the DenseNet-121 architecture and transfer learning. With an accuracy of 82.1%, the recommended model finished the classification task. In order to avoid bias towards a class with a greater picture count, the authors emphasise how important it is to employ balanced datasets in classification tasks. The study shows how promising current technology is for treating and diagnosing skin cancer. Taken together, the proposed approach provides a viable approach to accurate skin lesion classification. In ”Deep Learning-Based Classification of Dermoscopic Images for Skin Lesions” by Ahmet Furkan Sonmez et al., [2] a deep learning-based approach using the MobileNetV2 model is pro- posed to classify skin lesions from dermoscopic images, achieving an accuracy of 80.79% on the HAM10000 dataset. The dataset consist saven classes. The study aims to demonstrate the potential of deep learning in accurately categorizing skin lesions, aiding in early detection and treatment of skin cancer. However, the focus on a specific dataset limits the generalizability of the proposed method to diverse clinical settings and populations. Despite this limitation, the research contributes to advancing the application of deep learning in dermatology, paving the way for future developments in computer-aided diagnosis and personalized healthcare. An innovative method of classifying skin lesions using a pre- trained DenseNet201 architecture is presented in the publication ”Skin Lesion Classification Using a Pre-Trained