Page 49 - AIH-2-4

Artificial Intelligence in Health ViT for neurodegeneration diagnosis

Table 4. The performance comparison of our model with others in the literature
Model Class Sensitivity Specificity Precision F1 score F1 score (micro‑average)
Ding et al. (model CNN) CN 0.59 0.75 0.60 0.59 0.64
5
MCI 0.54 0.68 0.55 0.55
AD 0.81 0.94 a 0.76 0.78
Etminani et al. (model CNN)* CN 0.88 0.90 0.81 0.84 0.63
4
MCI 0.17 0.94 a 0.20 0.18
AD 0.91 a 0.92 0.83 a 0.87 a
Etminani et al. (consensus of human readers)* † CN 0.70 0.81 0.64 0.67 0.45
4
MCI 0.25 0.75 0.08 0.12
AD 0.47 0.90 0.68 0.56
Our model (ViT) CN 0.90 a 0.95 a 0.90 a 0.90 a 0.81 a
MCI 0.65 a 0.90 0.76 a 0.70 a
AD 0.90 0.88 0.78 0.84
Notes: All results were obtained using F-FDG PET brain scans from the ADNI dataset. Besides acquiring a significantly higher F1 score, our model
18
outperforms others in classifying MCI cases by a considerable margin. *The authors considered an additional DLB class in their paper and tested the
model and human readers in a four-class classification task (CN, MCI, AD, DLB). Four professional nuclear medicine physicians with 3, 8, 13, and 16
†
years of experience. The highest values per each metric and class.
a
Abbreviations: AD: Alzheimer’s disease; ADNI: Alzheimer’s Disease Neuroimaging Initiative; CN: Cognitively normal; CNN: Convolutional neural network;
DLB: Dementia with Lewy bodies; F-FDG: F-fluorodeoxyglucose; MCI: Mild cognitive impairment; PET: Positron emission tomography; ViT: Vision
18
18
transformer.
impact on its decision and a heatmap of the input scan, in applicable models that can assist experts in NDD diagnosis.
which the pixels’ intensities illustrate their importance in
the model’s prediction. Furthermore, we conducted a study 6. Conclusion
to reveal regions’ significance in the model’s decisions, We believe this research showcases the extraordinary
showing some brain areas are of utmost importance in potential of the ViT architecture in NDDs classification,
predicting various conditions. which surpasses other methods, including CNNs. Apart
from their excellent performance, ViTs allow computer
5.1. Limitations
vision researchers to benefit from advancements in NLP
A key goal of this study was to ensure that the model can due to the sharing of the same transformer architecture.
distinguish and classify all stages of AD. Therefore, we Furthermore, ViTs make developing explainable models
decided to develop the model solely on MCI cases that more feasible by leveraging the attention mechanism.
later progressed to AD. However, this sample selection
might introduce some bias in the dataset regarding the Acknowledgments
MCI class and result in the model becoming a prognostic None.
MCI-to-AD classifier. In addition, we exclusively relied
on the ADNI dataset in this research, which may restrict Funding
the model’s out-of-distribution generalization. Data collection and sharing for this project was funded
Although the critical regions suggested by the proposed by the ADNI (National Institutes of Health Grant U01
model are consistent with the literature to a significant AG024904) and DOD ADNI (Department of Defense
extent, our findings require more examination and award number W81XWH-12-2-0012). ADNI is funded
validation by medical domain experts. by the National Institute on Aging, the National Institute
of Biomedical Imaging and Bioengineering, and through
Finally, we should stress that while DNNs demonstrate generous contributions from the following: AbbVie,
promising results in NDD classification, they are limited to Alzheimer’s Association; Alzheimer’s Drug Discovery
their datasets, substantially affecting their generalization Foundation; Araclon Biotech; BioClinica, Inc.; Biogen;
performance. In addition, unlike clinical procedures, DNNs Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate;
do not consider all medical factors and base their predictions Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and
on limited data. Therefore, vast improvements and training on Company; EuroImmun; F. Hoffmann-La Roche Ltd and
diverse datasets are critical to designing robust and clinically
its affiliated company Dec 5, 2024 12:30 PM Genentech,

Volume 2 Issue 4 (2025) 43 doi: 10.36922/AIH025140026