Reducing Labeling Costs in Alzheimer's Disease Diagnosis: A Study of Semi-Supervised and Active Learning with 3D Medical Imaging

• Published in: 2023 International Conference on Modeling, Simulation & Intelligent Computing (MoSICom) pp. 264 - 269
• Main Authors: Qayyum Patel, Rida Abdul, Mihailescu, Radu-Casian
• Format: Conference Proceeding
• Language: English
• Published: IEEE 07.12.2023
• Subjects: active learning; alzheimer's disease; Annotations; Brain modeling; convolutional neural network; Costs; deep learning; Magnetic resonance imaging; Robustness; semi-supervised active learning; semi-supervised learning; Three-dimensional displays; Training
• DOI: 10.1109/MoSICom59118.2023.10458754

Alzheimer's Disease (AD) is a neurodegenerative condition that is the most common cause of dementia. While there is no cure, its early detection is crucial for effective medical intervention. Deep learning models trained on brain Magnetic Resonance Imaging (MRI) scans have shown promise in this regard, but obtaining annotations for medical imaging data is expensive. In this study, we explore three network training approaches that aim to minimize labeling costs - Active Learning (AL), Semi-Supervised Learning (SSL), and Semi-Supervised Active Learning (SSAL). These were applied to train a 3D subject-level convolutional neural network to diagnose AD using 3D brain MRI scans. Our results confirm the significant impact of the annotation budget and the initial training set on model performance. We observe that all approaches consistently outperform random sampling. Uncertainty-based AL achieves comparable performance to the traditional supervised baseline using only 30 percent of the annotated data. Representative AL and joint SSAL outperform the traditional supervised baseline using 30 percent of the annotated data, with the latter showing robustness even with a restricted initial training set.