A grad-CAM and deep learning based image classification for skin diseases

No Thumbnail Available

Date

2022-07

Journal Title

Journal ISSN

Volume Title

Publisher

International Islamic University Chittagong

Abstract

Skin disease classification and detection are critical and cutting-edge research areas within the field of DeepLearning. With the objective of enhancing the performance of existing di agnostic systems, this study explores the application of deep convolutional neural networks (CNNs), leveraging advanced architectures and visualization techniques. In addressing these challenges, this paper focuses on implementing a customized CNN model while incorporating ResNet-50 and VGG-16 to leverage their combined strengths for enhanced diagnostic accuracy. Additionally, the study employs Grad-CAM for effective visualization of model predictions and decision-making processes.Utilizing the ISIC 2019-2020 dataset, which comprises 9,200 images, our approach encompasses a comprehensive training and testing regimen. The dataset’s diver sity and volume provide a robust foundation for evaluating the efficacy of our proposed models. In parallel, the ISIC Archive dataset further enriches our experimental framework, enabling a nuanced assessment of model performance across varied data scenarios. The custom CNN model attained an accuracy of 87% on the ISIC 2019-2020 dataset, while the CNN+ResNet and CNN+VGG16 configurations achieved accuracies of 95% and 92%, respectively. Regarding the ISIC Archive dataset, the custom CNN model demonstrated a 91.91% accuracy, while the CNN+ResNet and CNN+VGG16 achieved 92% and 94%, respectively.Our findings contribute a novel perspective to the domain of skin disease detection, offering a scalable and efficient frame work that outperforms existing methodologies. The integration of Grad-CAM into our model architecture underscores our commitment to advancing both the accuracy and interpretability of deep learning-based diagnostic systems in dermatology.

Description

This Dissertation is Submitted in Fulfillment of the Requirements for the Degree of Bachelor of Science (B.Sc.)

Keywords

Deep Learning, ,VGG16, ResNet-50, Grad-CAM.

Citation