COMPARATIVE ANALYSIS OF NEURAL NETWORK ARCHITECTURES FOR MEDICAL IMAGE CLASSIFICATION: EVALUATING PERFORMANCE ACROSS DIVERSE MODELS

Abstract

Medical image classification has become a critical task in computer-aided diagnosis, enabling faster and more accurate detection of diseases such as cancer, pneumonia, and diabetic retinopathy. This study presents a comprehensive comparative analysis of prominent neural network architectures—Convolutional Neural Networks (CNNs), Residual Networks (ResNets), DenseNets, Vision Transformers (ViTs), and EfficientNets—in the context of medical image classification. Utilizing benchmark datasets including ChestX-ray14, ISIC Skin Cancer, and Retinal Fundus Images, we evaluated each model's performance based on accuracy, precision, recall, F1-score, training efficiency, and robustness to overfitting. The results demonstrate that while CNN-based models like ResNet and DenseNet maintain strong classification capabilities with balanced computation cost, ViTs outperform others in high-resolution image interpretation, especially under complex feature distributions. EfficientNet offers a trade-off between speed and accuracy, making it suitable for resource-constrained clinical settings. Our findings highlight the architectural strengths and weaknesses in varied medical imaging scenarios, providing insights into the selection of optimal models based on diagnostic goals, dataset characteristics, and computational resources..