Improving Performance of VGG-19 Model using Dual Input Block for Skin Disease Classification

Abstract

Skin diseases are a growing threat to humans. They can cause severe damage to the skin. The damage to the skin sometimes can lead the patient to have a loss of confidence and depression. By using emerging technologies like deep learning, we can identify skin diseases and treat them effectively. These computer-powered techniques can detect skin diseases without any help from a professional. People can have a user-friendly experience if some effort is put into the interface. This can help people identify the disease in its early stages instead of just ignoring it as an allergy. This will help people avoid suffering from the severe consequences of the diseases. By using these technologies, we can accurately find the disease in the patient. By using deep learning algorithms like those at CNN, we can make things simpler and less time-consuming. By using advanced models like Inception Net and ResNet and optimizing the model's hyperparameters like learning rate, optimizers, etc., the model can give high performance and accuracy. The model can be trained to give accurate results even if the images used in the training dataset and the image given by the patient are of low quality. We propose a dual-input CNN model to identify skin diseases. We use a modified VGG-19 neural network with dual input blocks and batch normalization to classify the diseases. The model outperforms the original VGG19 by about 5 percent. The original model, trained from scratch, was also modified to have batch normalization. The models did not learn, and the accuracy did not increase without batch normalization. The models were trained for 150 epochs. Both were trained under the same parameters and methods. The dual input model yields about 94%, while the original VGG19 architecture trained from scratch gives about 89%. The models were trained on P100 GPU are available in Kaggle. The model is implemented on hardware that can give results without any delay to the patient.