Transfer learning-based malnutrition classification using VGG16 and comparative analysis of CNN architectures

Abstract

Childhood malnutrition remains a critical public health challenge in developing countries, with Indonesia ranking fifth globally for stunting burden. Traditional anthropometric assessment methods are time-consuming, resource-intensive, and require trained personnel, necessitating efficient computer-based early detection approaches. This study proposes a deep learning-based method for automated nutritional status classification using facial image analysis. We developed and compared multiple transfer learning architectures including visual geometry group 16 (VGG16), densely connected convolutional network 121 (DenseNet121), mobile network version 2 (MobileNetV2), and residual network 50 (ResNet50) for classifying children into three categories: healthy, malnutrition, and stunting. Results demonstrated that VGG16, a simpler architecture trained for only 10 epochs, achieved optimal performance with 91.8% accuracy, and significantly outperforming more complex modern architectures like ResNet50. This finding challenges the conventional assumption that newer, deeper models invariably perform better, and particularly when working with limited medical datasets. The study revealed that longer training durations led to performance degradation due to overfitting, emphasizing the importance of balancing model complexity with dataset characteristics. These findings support the development of practical artificial intelligence (AI)-based malnutrition screening systems suitable for resource-constrained environments, potentially improving early detection capabilities, and public health outcomes in developing regions.