Publications
List of research works done in the past and ongoing in reverse chronological fashion
DomainAdapt: Leveraging Multitask Learning and Domain Insights for Children’s Nutritional Status Assessment
In Press, Link to dataset
Medical Image Computing and Computer Assisted Intervention (MICCAI) 2024 (Accepted)
In this research, we propose DomainAdapt, a novel multitask learning framework designed to enhance nutritional status assessment for children using a combination of domain knowledge and Mutual Information. By dynamically adjusting task weights based on domain insights, DomainAdapt achieves superior accuracy in predicting anthropometric measures such as height, weight, and Mid-Upper Arm Circumference (MUAC), as well as classifying malnutrition-related conditions like stunting, wasting, and underweight.
A key contribution of our study is the introduction of AnthroVision, a new and comprehensive dataset consisting of 16,938 multipose images from 2,141 children, capturing diverse backgrounds, clothing, and lighting conditions across clinical and community settings. AnthroVision provides a rich source of data for automating nutritional assessments, enabling more accurate and scalable solutions compared to traditional methods. Through rigorous experimentation, our proposed model demonstrates significant improvements over existing multitask learning approaches, offering a robust framework for public health applications in resource-constrained environments.
NutriAI: AI-Powered Child Malnutrition Assessment in Low-Resource Environments
In Proceedings of the 32nd International Joint Conference on Artificial Intelligence (IJCAI-23), Special Track on AI and Social Good, 19th-25th August 2023, Macao, S.A.R.
Malnutrition remains a pressing global health challenge, particularly in low- and middle-income countries where resource limitations make traditional assessment methods difficult to implement. NutriAI is a project proposal that aims to introduce an AI-driven solution leveraging computer vision and deep learning to assess malnutrition in children through 2D images. By providing a low-cost, scalable alternative to traditional anthropometric measurements, NutriAI addresses the challenges of manual assessments, which can be time-consuming and prone to inconsistencies.
The proposed solution utilizes a dual-branch framework that analyzes facial and full-body images, estimating anthropometric indicators such as height, weight, and middle-upper-arm circumference (MUAC). These are then compared with established health benchmarks from the World Health Organization (WHO) and Centers for Disease Control and Prevention (CDC) to detect malnutrition. This approach is designed to improve the accuracy and efficiency of malnutrition screening, making it suitable for large-scale use in low-resource settings.
​
On AI Approaches for Promoting Maternal and Neonatal Health in Low Resource Settings: A Review
Frontiers in Public Health
The article discusses challenges in low- and middle-income countries for timely medical diagnosis of maternal and neonatal diseases. It explores the potential of digital technologies and artificial intelligence as practical assistive tools in maternal and neonatal healthcare. It also highlights open problems and future research directions in this area.
Novel Few Shot Learning Techniques for Nutritional Status Detection via Images
Women in Computer Vision (WiCV) at CVPR 2022 (A*)
Accepted - Poster, Travel Award
This paper explores using computer vision for nutritional status detection in low-income and middle-income countries. The authors propose MalDB, a database of annotated images of malnourished and healthy children. They conducted experiments using deep learning models achieving up to 86.94% accuracy using ensemble approaches. The study aims to improve the accessibility and accuracy of nutritional status detection in resource-constrained settings.