Predicting osteoporosis using fundus images

• Main Author: Loke, Brandon Ming Kai
• Other Authors: Kwoh Chee Keong
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Osteoporosis prediction; Fundus images; Machine learning; Image classification
• Online Access: https://hdl.handle.net/10356/175384
• Institution: Nanyang Technological University

Osteoporosis is a prevalent skeletal disorder characterized by reduced bone mineral density levels which presents significant health challenges. While prior studies have explored predicting the presence of this disease using machine learning techniques, they have primarily relied on specific bone images or DEXA scans, which can be challenging to obtain due to the cost of the equipment as well as the procedure itself being relatively troublesome. This study investigates the feasibility of predicting osteoporosis using fundus images, a noninvasive and accessible modality. In collaboration with the National Healthcare Group, a dataset consisting of 3450 entries of fundus images and corresponding tabular data was leveraged for model training. The objective is to develop a predictive model capable of accurately classifying patient data into three categories: Normal Bone Density, Low Bone Density, and Osteoporosis. This study employs ResNet50 and InceptionV3 CNN models, fine-tunes their hyperparameters, and integrates tabular data using a joint fusion (type-1) model. Additionally, ensemble learning via soft voting is employed for prediction aggregation. Performance evaluation metrics include accuracy, F1-score, precision, recall, as well as the ROC-AUC values for each class. Results demonstrate promising performance, with the ResNet50 joint fusion (Type-1) model achieving the highest accuracy and precision of 88.53% and 90.03%. While the individual ResNet50 model achieved the highest F1-score of 88.83%, and highest recall of 87.25%. This study will allow us to efficiently identify individuals at risk of osteoporosis, enabling timely intervention and management strategies