Abstract
The existing diagnostic methods for Alzheimer’s disease (AD) require considerable time and expense from patients, and approaches employing machine learning have also failed to address these issues. We propose a mobile device-targeted AD diagnosis network, which uses retinal fundus images. The proposed model adopts MobileNetV3-Large as a backbone for efficient computation with a structure modified based on the UNet architecture. We used an attention mechanism to obtain higher diagnosis performance. To improve the image quality robustness of our model, training was conducted using input image masking with the strategy of random erasing data augmentation method. We demonstrate that our proposed model performs superior to other mobile device-targeted state-of-the-art (SOTA) models by achieving an area under the receiver operating characteristic curve (AUC) of 0.927 on the validation dataset from the UK Biobank. Finally, we also successfully emulated the proposed model and tested its operation on a smartphone application. We can confirm that the proposed method can reduce the utilization of these resources while achieving higher diagnosis performance. This research demonstrates that the proposed approach can be used as a clinical AD diagnosis assistance tool in our smartphones.