DEEP LEARNING-BASED EARLY DETECTION OF MALNUTRITION IN CHILDREN UNDER FIVE YEARS

Abstract

Malnutrition remains a critical public health concern, particularly among children, necessitating early and accurate diagnostic methods to support timely intervention. This project presents a deep learning-based system for the automatic detection of malnutrition in children using full-body pediatric images. The system classifies images into two categories—Malnutrition and Nutrition—leveraging Convolutional Neural Network architectures. The application is developed using Python for backend processing and Flask as the web framework, with a responsive interface built using HTML, CSS, and JavaScript. Two state-of-the-art deep learning models, ResNet152V2 and MobileNet, were trained and evaluated independently on a balanced dataset comprising 2,010 images (1,005 for each class). The image preprocessing pipeline includes rescaling, resizing, data augmentation, directory-based loading, batching, shuffling, and training-validation splitting, ensuring robust training and model generalization. The ResNet152V2 model achieved a training accuracy of 99% and test accuracy of 96%, while the MobileNet model recorded a training accuracy of 99% and test accuracy of 95%. For performance evaluation, standard classification metrics such as Precision, Recall, F1-score, and Confusion Matrix were utilized. Additionally, model accuracy and loss graphs were plotted to visualize training performance across epochs for both models. The results demonstrate that deep learning can effectively distinguish between malnourished and nourished children using visual cues alone, offering a scalable and non-invasive approach to nutritional screening in clinical and rural settings. This system has the potential to support healthcare professionals and aid in the early detection of malnutrition, thereby improving pediatric health outcomes