Empirical Study on Modified Pre-Trained CNN Architectures for Fitzpatrick17k Skin Diseases Prediction Modelling

Abstract

Preventing missed diagnosis of skin diseases is critical for enhancing patient outcomes. Technology support empowered by AI based algorithms have been recently developed to minimize the burden. Convolutional Neural Networks (CNNs) are increasingly utilized for image classification tasks including skin's diseases prediction modelling. Its advancement architectures such as pre-trained models, VGG16 andResNet50 provide a strong foundation to up-skill the prediction capability. However, it takes higher resources. This research proposed empirical study of pre-trained modified CNN's architecture in handling skin's diseases identification which include the challenging open skin's diseases dataset, Fitzpatrick 17k as the task. The proposed modified CNN is pre-trained by using large skin diseases dataset ISIC 2019. The study involves the performance evaluation of pre-trained model and modified CNN architecture for classifying skin diseases using the Fitzpatrick17k dataset, which includes a diverse representation of skin tones and conditions. Simulation’s findings demonstrate that the proposed pre-trained modified CNN architecture improved performance by up to 6% in accuracy compared to the baseline VGG16 model. While freezing specific layers enhanced model accuracy, this approach introduced trade-offs, such as a decrease in precision, which should be addressed in future research to optimize overall model performance. In addition, this research elaborates discussion on overfitting issue in handling Fitzpatrick17k-C dataset.