BRIDGING LAB-TO-CLINIC DEPLOYMENT: EXPLAINABLE, OFFLINE DEEP LEARNING FOR ALZHEIMER’S MRI ANALYSIS ON MOBILE DEVICES

Abstract

Alzheimer’s disease is by far one of the most challenging neurodegenerative disorders that is out there, and its early detection is really game-changing as its early detection can be handled very well. As of now, many doctors mainly rely on clinical exams, such as diagnosis through MRIs, but such tools are very costly and are very time-consuming, and their availability is another issue, especially in remote and rural areas. In research, deep learning models such as Convolutional Neural Networks (CNNs) do a really great job by picking up signs of Alzheimer’s disease from MRI images [1]. But the problem is that most of these models are usually cloud-based, meaning patient data gets sent online, raising plenty of red flags, such as privacy, speed, and whether these AI tools fit into clinics or not. Also, they will be available in case of no internet. Due to these reasons, deep learning isn’t thought to be a quick and on-the-spot screening tool by now [16]. So, in this paper, we introduce Deep Cortex-Ai, which is a complete framework that is designed to make Alzheimer’s screening closer to real-world use. We have made an Android-friendly app that runs offline, making it as accessible as possible and privacy-friendly. At the core is the Xception architecture, which we fine-tuned using a two-phased learning process using transfer learning by training the model on 27,188 brain MRI images (37,386 total) to make sure it generalizes well. The system accurately sorts scans into four categories: Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented. What really sets our approach apart is that we put the trained model directly onto mobile devices with TensorFlow Lite. Users can get results in real time and don’t need an internet connection, as the data never leaves the device unless the user shares it. Deep Cortex showed an accuracy of 89.59% on more than 3 thousand test images. For the cases of non demented, it showed a near-perfect accuracy score. The app itself is lightweight weight working on ordinary mid-range devices, with analysis taking less time than

3 seconds on average. For clinicians to understand and trust the app's decisions, we have integrated Grad-CAM visualizations, which highlight the neuroanatomical regions, hippocampus, and ventricles that matter the most for the model’s predictions. The interface and the workflow are also easy for the clinicians to get used to: easy MRI upload, clear probability scores, and automatic,

detailed PDF reports that are ready to share. With this approach, we are moving a step forward for easy and early screening of Alzheimer’s that is accessible everywhere. Opening the doors to early intervention and a positive outcome for the patients around the world.