A FixMatch Framework for Alzheimer’s Disease Classification: Exploring the Trade-Off Between Supervision and Performance
Alzheimer’s Disease (AD) poses a major challenge for healthcare systems worldwide, as timely and accurate diagnosis is crucial for patient management and outcome improvement. While experienced medical professionals can often identify AD through conventional assessment methods, limited res...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10946883/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Alzheimer’s Disease (AD) poses a major challenge for healthcare systems worldwide, as timely and accurate diagnosis is crucial for patient management and outcome improvement. While experienced medical professionals can often identify AD through conventional assessment methods, limited resources and growing patient populations make large-scale and rapid screening increasingly necessary. In this work, we explore whether the FixMatch algorithm—a semi-supervised learning approach—can aid in classifying Alzheimer’s Disease (AD), Mild Cognitive Impairment (MCI), and Cognitively Normal (CN) by using the ADNI fMRI dataset of 5,182 images. This approach supplements rather than replaces clinical expertise, offering a faster, more standardized classification process where expert labeling is limited. We first assessed various supervised models and determined that VGG19-FFT provided the strongest balance of classification accuracy and computational efficiency. Integrating VGG19-FFT into FixMatch as the teacher model, initial tests using 10% labeled and 90% unlabeled data yielded modest results. A more systematic examination of different labeled-to-unlabeled data splits revealed that a 60:40 ratio enabled FixMatch to achieve classification accuracies of 100% for AD, 99% for CN, and 99% for MCI—on par with fully supervised training. This outcome highlights the potential of FixMatch to significantly reduce labeling requirements, a particular advantage in resource-constrained settings where expert annotations are costly. By striking an effective balance between labeling effort and model performance, the identified 60:40 ratio helps make advanced diagnostic methods both feasible and practical in real-world clinical applications. |
|---|---|
| ISSN: | 2169-3536 |