Optimizing Skin Disease Classification: Evaluating the Effectiveness of Hybrid CNN with Batch Normalization and L2 Regularization for Enhanced Accuracy

Abstract

Skin disease classification remains a significant challenge in medical image analysis, demanding robust models for accurate diagnosis. In this research paper, we meticulously explore and compare the performance of various neural network architectures on a curated skin disease dataset. Our study includes traditional models such as Multi-Layer Perceptron (MLPs) with different hidden layers, a powerful deep architecture like ResNet and proposed hybrid Convolutional Neural Networks (CNNs) with Batch Normalization and L2 Regularization. MLP classifiers, equipped with one and two hidden layers, demonstrate promising F1 scores, sensitivity, and specificity, showcasing their effectiveness in skin disease classification tasks. The Sequential Neural Network, a versatile architecture, exhibits commendable accuracy, precision, recall, and F1 score, highlighting its suitability for medical image analysis. Our standout performer is the proposed hybrid CNN model with Batch Normalization and L2 Regularization, achieving an impressive accuracy of 97%. This hybrid architecture outclasses other classifiers, emphasizing the impact of advanced techniques in enhancing model performance. By using various assessment criteria and diagrams, the suggested paper offers crucial information about the advantages and disadvantages of each model, helping researchers and practitioners choose the optimal architecture for skin disease classification. These findings augment the ongoing study in medical image analysis using deep learning techniques and reemphasizes the effectiveness of model selection in diagnosing diseases accurately and comprehensively.