Decision Support System for Detection and Classification of Skin Cancer using CNN

Decision Support System for Skin Cancer Detection and Classification using CNN

Introduction

Skin cancer, a predominant public health concern, necessitates early detection and classification to mitigate its spread and improve patient outcomes. Leveraging Convolutional Neural Networks (CNN) and Transfer Learning, this study outlines a decision support system aimed at accurate skin cancer classification using the MNIST HAM-10000 dermoscopy image dataset. The growing prevalence of skin cancer and the challenges inherent in early-stage detection underscore the significance of this research.

Literature Review

The utility of CNNs in medical imaging, particularly in skin lesion analysis, has been well-documented. Prior works have showcased CNNs' superiority over traditional methodologies and even trained medical professionals in certain instances. The integration of Transfer Learning, utilizing pre-trained models like VGG-16 on extensive datasets such as ImageNet, has further amplified the accuracy rates of skin cancer classification systems. Despite these advancements, existing literature predominantly focuses on binary classification (cancerous vs non-cancerous), with multi-class classification of skin lesions remaining relatively underexplored.

Methodology

The study employs the MNIST HAM-10000 dataset, comprising over 10,000 skin lesion images across seven distinct categories. This dataset serves as the cornerstone for the CNN model development and subsequent Transfer Learning application. Initial preprocessing steps include data augmentation and histogram equalization techniques to address class imbalance and enhance image quality. The methodology pivots on a CNN framework complemented by Transfer Learning to refine the model's classification accuracy.

Convolution Neural Network (CNN) Model

• The CNN architecture is built upon layers designed to mimic the hierarchical pattern recognition of the human visual cortex. Key components include convolutional layers, max-pooling, and fully-connected layers facilitating feature extraction and classification.

Transfer Learning Approach

• Leveraging pre-trained models such as VGG16 and ResNet, Transfer Learning adapts these architectures to the domain-specific task of skin cancer classification. This strategy not only expedites the training process but also enhances model performance by utilizing the vast, generic feature sets learned from the ImageNet dataset.

Results

The implementation revealed promising outcomes, with the CNN model achieving a weighted average precision of 0.88, recall of 0.74, and an F1-score of 0.77. The application of Transfer Learning, particularly using the ResNet model pre-trained on ImageNet, resulted in a notable accuracy increase to 90.51%. Comparative analysis with other machine learning algorithms, including Random Forest, XGBoost, and SVM, underscored CNN and Transfer Learning's superior performance for this task.

Conclusion

This study underscores the effectiveness of CNNs and Transfer Learning in enhancing the accuracy of skin cancer classification systems. By adopting a Transfer Learning approach with pre-trained models, the research significantly improved the classification accuracy of skin lesions depicted in the HAM10000 dataset. Future directions include refining the model to elevate prediction outcomes and further exploring the utility of Transfer Learning across additional datasets and classification frameworks.

The implications of such advancements extend beyond academic interest, offering tangible improvements in diagnostic methodologies and patient care strategies in dermatology. As AI continues to evolve, its integration into healthcare, particularly in areas requiring nuanced interpretation of medical images, heralds a new era of precision medicine.