Skin Lesion Diagnosis Using Deep Neural Networks
Subject Areas : تخصصی
Mahdi Hariri
1
,
ُSoudabeh barzegari
2
1 - Assistant Professor, Electrical engineering Department, Faculty of Technical and Engineering, University of Zanjan, Zanjan, Iran
2 - Master's student, Department of Electrical Engineering, Faculty of Engineering, University of Zanjan, Zanjan, Iran
Keywords: Skin cancer, melanoma, convolution, segmentation, deep learning,
Abstract :
Skin cancer is one of the most common cancers in the world. Early detection of cancerous lesions is of great importance in treatment. Skin lesion images contain important information for classification, and due to the diversity of lesion shapes, automated image processing systems effectively help in diagnosing the type of lesion. Due to the appropriate accuracy of artificial intelligence, especially deep learning methods in image classification, their use in medical image classification is also expanding. Despite their appropriate accuracy, these models have a large computational burden that limits their use. The use of lighter deep learning algorithms increases the hope of using them as applications on mobile phones in society.
In this study, an efficient model for classifying skin lesions has been proposed to help diagnose the disease. In this model, four convolutional layers, two merging layers, and two batch normalization layers were used. This model helps to identify the correct class of input samples by structurally examining similarities and has been tested on images of a wide range of skin cancer types from different individuals. The skin lesions in this set are distributed in seven main classes. Using the technique of increasing the number of samples, we correct the imbalance of the dataset used. In classifying the data set by the presented model, the accuracy and precision of the proposed method were 87.72% and 89.1%, which due to the number of parameters and smaller volume, the proposed method has improved compared to the group learning methods, simple convolutional network, and transfer learning