Implementation of MobileNet Architecture for Skin Cancer Disease Classification

Abstract

As the number of occurrences of skin cancer increases year, it becomes more and more crucial to identify the disease accurately and effectively. This study aims to implement and evaluate the MobileNet architecture for classifying nine types of skin lesions using the ISIC 2020 dataset and to compare MobileNet's performance with other CNN architectures, such as VGG-16 and LeNet, in terms of accuracy and computational efficiency. The study also investigates the impact of image preprocessing on model accuracy. The methodology comprises data collection, preprocessing, and model development, leveraging transfer learning from MobileNet pre-trained on ImageNet. Data preprocessing involves resizing images to 224 x 224 pixels and normalizing pixel values. To augment the dataset, techniques such as rotation, zooming, horizontal flipping, and brightness and contrast adjustment are applied. To address class imbalance, oversampling is used to obtain 500 images per class. The model architecture includes Global Average Pooling, a Dense layer with 1024 units and ReLU activation, and a Dropout layer with a 0.2 probability. Various training scenarios with batch sizes (8, 16, 32, 64) and learning rates (0.001, 0.0001) are evaluated, incorporating dropout and ReLU activations. The optimal performance was achieved with oversampling, dropout, and a learning rate of 0.0001, yielding a training accuracy of 99.64% and a validation accuracy of 86.89% after oversampling, resulting in 3,600 training and 900 validation images with an 80:20 data split. The results suggest overfitting due to dataset limitations. Future work should focus on fine-tuning and ensemble methods to improve validation performance.