Hierarchical approach for the classification of multi-class skin lesions based on deep convolutional neural networks Samia Benyahia 1, Boudjelal Meftah 12 , Olivier LezoraY 3 ICPRAI 2022 - 3rd International Conference on Pattern Recognition and Artificial Intelligence , June 2022 1 Department of Computer Science, Faculty of Exact Sciences, University of Mascara, Mascara, Algeria 2 LRSBG Laboratory, University of Mascara, Mascara, Algeria 3 Normandie Univ, UNICAEN, ENSICAEN, CNRS, GREYC, Caen, France

Introduction • A skin lesion refers to any skin area with different characteristics from the surrounding skin, including color, shape, size, and texture. • Skin cancer arises from the uncontrolled growth and spread of abnormal cells that grow beyond their usual boundaries. • Two main types of skin cancer are keratinocyte carcinoma and melanoma. • Each type of skin cancer has unique characteristics. • Skin cancer is difficult to be diagnosed, as malignant skin lesions can closely resemble their benign counterparts. • Different lesion types can have similar characteristics, furthering the problem of discriminating among them. • 2

• Variety of deep learning methods were proposed to diagnose dermoscopy images. • Deep learning techniques do not understand the information they are dealing with, they simply try to detect patterns or correlations among the different skin lesions. • The hierarchical organization of lesion types defined by dermatologists which could be used to provide a better understanding of the diagnosis made by the system. 3 Introduction

Methodology 4 Fig.1. Proposed method's fowchart Skin lesion Dataset Train test CNN CNN CNN CNN CNN MEL NV SCC BCC BKL VASC AK DF

Convolutional neural networks 5 • Convolution, pooling layers, transition layers, dense blocks, fully connected layers • Dense blocks • DensNet-121, DensNet-169,DensNet-201, and DensNet-246 ResNet DenseNet • Convolutional, maxpool, batch normalization, activation, and a residual blocks, fully connected layers • Skip-connection between layers • ResNet18, ResNet34, ResNet50, ResNet101, and ResNet152

Experiments and Results Two scenarios : • Sequential • Hierarchical classification approache 6 Table 1. results obtained with sequential and hierarchical Resnet50 and DenseNet201 architectures. 0,00% 20,00% 40,00% 60,00% 80,00% 100,00% 120,00% Accuracy Sensitivity Speci city Precision Sequential ResNet50 Hierarchical ResNet50 Sequential DenseNet201 Hierarchical DenseNet201

Experiments and Results • Unbalanced dataset • Data augmentation • Using upsampling and downsampling 7 0,00% 20,00% 40,00% 60,00% 80,00% 100,00% 120,00% Accuracy Sensitivity Speci city Precision Sequential ResNet50 Hierarchical ResNet50 Sequential DenseNet201 Hierarchical DenseNet201 Table 3. results obtained with sequential and hierarchical ResNet50 and DenseNet201 architectures with data augmentation.

Conclusion • A hierarchical CNN-based system for classifying multi-class skin lesions are proposed. • comparative studies between the proposed hierarchical model and a sequential model using different CNN architectures: Resnet50 and DenseNet201. • The obtained results highlight the benefits of addressing the classification of different skin lesions with CNNs in such a hierarchically structured manner. • Data augmentation considerably improves the results of classification for both hierarchical and a sequential model. 8

References 1. Skin cancer statistics from skin cancer foundation. https://www.skincancer.org/ skin-cancer-information/skin-cancer-facts/. Accessed November 2020 2. https://www.ncbi.nlm.nih.gov/books/NBK247164/figure/skincancer.f3/ 3. Howlader N, Noone AM, Krapcho M, et al. (eds). SEER Cancer Statistics Review, 1975–2018. National Cancer Institute, posted to the SEER website, April 2021. Last accessed April 19, 2021 4. He, K., Zhang, X., Ren, S., Sun, J.:Deep Residual Learning for Image Recognition,IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770-778(2016) 5. Huang, G., Liu, Z.,Weinberger, K. Q.:Densely Connected Convolutional Networks, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2261-2269 (2017) 6. Tschandl P., Rosendahl C. & Kittler H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Sci. Data 5, 180161 doi.10.1038/sdata.2018.161 (2018) 7. Noel C. F. Codella, David Gutman, M. Emre Celebi, Brian Helba, Michael A. Marchetti, Stephen W. Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, Allan Halpern: "Skin Lesion Analysis Toward Melanoma Detection: A Challenge at the 2017 International Symposium on Biomedical Imaging (ISBI), Hosted by the International Skin Imaging Collaboration (ISIC)", 2017; arXiv:1710.05006. 8. Marc Combalia, Noel C. F. Codella, Veronica Rotemberg, Brian Helba, Veronica Vilaplana, Ofer Reiter, Allan C. Halpern, Susana Puig, Josep Malvehy: "BCN20000: Dermoscopic Lesions in the Wild", 2019; arXiv:1908.02288. 9

Thanks! 10