A multitasking deep learning system for the detection of Alzheimer's and Parkinson's disease using MRI.

Abstract

A multi-task deep learning system for the detection of Alzheimer's and Parkinson's disease using MRI, consisting of: a ResNet18-based Convolutional Neural Network (CNN) module configured to extract spatial features from preprocessed T1-weighted MRI scans; a first classification head coupled to the CNN module and configured to predict the disease type from a group consisting of Alzheimer's disease (AD), Parkinson's disease (PD) and a healthy control group; a second classification head which is conditionally activated by the first classification head when Alzheimer's disease is predicted and is configured to classify the Alzheimer's stage into one of the following stages: very mild, mild or moderate dementia; a combined loss function module configured to integrate categorical cross-entropy loss for disease classification and weighted cross-entropy loss for Alzheimer stage classification; an interpretability engine configured to provide predictive explanations and including the following: a Grad-CAM module for class-specific spatial activation mapping, a SHAP module for global feature mapping at the voxel level and a LIME module for the local, surrogate-based explanation of individual MRI scan predictions.

Description

[0001] The present invention relates to medical imaging and artificial intelligence, in particular a deep learning-based system for the automated detection and staging of neurodegenerative diseases such as Alzheimer's and Parkinson's using MRI scans of the brain.

[0002] The diagnosis of neurodegenerative diseases such as Parkinson's disease (PD) and Alzheimer's disease (AD) remains a significant clinical challenge due to overlapping symptoms and the subtlety of brain changes in the early stages. Conventional structural magnetic resonance imaging (MRI), particularly T1- and T2-weighted sequences, has traditionally been used to rule out alternative causes of parkinsonism and dementia, but often appears normal in early disease stages, limiting its diagnostic utility. Advanced MRI techniques, including neuromelanin-sensitive MRI for PD and volumetric analyses for AD, have improved detection by revealing disease-specific tissue changes and atrophy patterns; however, consistent and accurate differentiation of disease subtypes and stages remains difficult.Furthermore, while machine learning and deep learning methods hold promise for improving diagnostic accuracy and prognosis using neuroimaging data, challenges remain in integrating these methods into clinical practice due to data complexity, class imbalances, and the need for interpretable models. Therefore, there is an urgent need for a unified, explainable, and resource-efficient AI-based diagnostic system capable of accurately classifying neurodegenerative diseases and disease progression using standardized MRI scans to enable early diagnosis and improved clinical decision-making.

[0003] To solve this problem, the present invention offers a multitasking deep learning system for the detection of Alzheimer's and Parkinson's disease using MRI.

[0004] The system offers a unified, multitasking deep learning system capable of detecting multiple neurodegenerative diseases - including Alzheimer's (AD), Parkinson's (PD) and healthy controls - using a single T1-weighted MRI-based framework.

[0005] The system offers a conditional Alzheimer's stage classification function that is only activated when AD is detected, thereby optimizing computational efficiency and ensuring clinically meaningful predictions.

[0006] The system integrates several explainable AI tools (XAI) - Grad-CAM, SHAP and LIME - into a single pipeline to generate spatial, global and local interpretability, promote physician confidence and support decision-making.

[0007] The system develops a lightweight, resource-efficient model architecture, such as a ResNet-18 based CNN, optimized for real-time inference on edge devices and use in resource-constrained clinical environments.

[0008] The system offers a modular and scalable architecture in which the disease classifier and the stage classifier can be operated independently or together, enabling seamless integration into clinical workflows, health information systems and cloud infrastructures.

[0009] The system corrects imbalances in the address class and variability between data sets through techniques such as weighted loss functions, data harmonization, and adaptive learning strategies, including learning rate schedulers and early stopping.

[0010] In one embodiment, the present invention provides a multitasking deep learning system for the detection of Alzheimer's and Parkinson's disease using MRI. The present invention discloses a novel, intelligent system for the automated diagnosis and staging of neurodegenerative diseases using structural brain imaging. More specifically, the invention relates to a unified, deep learning-based system that uses T1-weighted MRI scans of the brain to detect Alzheimer's disease (AD), Parkinson's disease (PD), and healthy controls, while conditionally classifying AD into clinically meaningful categories such as "very mild," "mild," or "moderate." The system is designed as a reliable, interpretable, and efficient tool to support clinical decision-making, facilitating the early detection and monitoring of neurological diseases.At the heart of the invention is a lightweight, multitasking learning system architecture based on a convolutional neural network (CNN) backbone (e.g., ResNet-18) and comprising two distinct classification modules: (a) a disease classification module that detects whether a scan indicates AD, PD, or a healthy brain, and (b) a conditional AD classification module that is activated only when the system predicts AD. This selective activation introduces task awareness and computational efficiency into the system, preventing unnecessary processing for non-AD cases and accurately mimicking clinical decision-making pathways. The system integrates three complementary interpretability tools—Grad-CAM, SHAP, and LIME—that work together to provide spatial, global, and local insights into the model's decision-making process.These interpretability features are crucial for clinical applications, as they allow healthcare professionals to visualize and validate which brain regions and features influenced the system's predictions. The system was specifically designed for ease of use and is therefore suitable for real-time operation in both high-performance computing environments and resource-constrained environments, such as rural clinics or edge computing devices (e.g., Raspberry Pi, NVIDIA Jetson). This design enables broad accessibility and scalability of the system beyond tertiary care centers. The invention will be explained again below.

[0011] The present invention discloses an intelligent assistive robotic system that supports elderly people and people with physical or medical limitations in maintaining their independence and safety during everyday activities. The system has a modular, malleable frame with joints driven by silent actuators, allowing it to physically adapt to various assistance functions such as mobility support, bedside assistance, or workstation support. It integrates sensors for real-time health monitoring, fall detection, and intuitive multimodal control interfaces, including voice, gesture, and touchscreen. A central AI engine continuously learns the user's preferences and routines and provides personalized support, medication reminders, and smart home control via wireless protocols such as Wi-Fi, Bluetooth, Zigbee, or Z-Wave.The system was developed with consideration for security, scalability and compliance with legal regulations, and features encrypted communication, emergency protocols and compatibility with ISO and IEC standards.

[0012] Furthermore, the invention comprises a deep learning-based diagnostic system for classifying and grading neurodegenerative diseases—in particular Alzheimer's disease (AD) and Parkinson's disease (PD)—using structural MRI data. Employing a lightweight dual-head CNN based on ResNet18, the system first classifies the disease type and then, if AD is detected, performs a staging into the stages "very mild," "mild," or "moderate." The architecture integrates several explainability modules (Grad-CAM, SHAP, LIME) to ensure clinical transparency and confidence. The system is optimized for deployment efficiency and flexibility and supports use on edge devices, in hospital networks, or on cloud platforms.Its modular and conditional design improves computing power while maintaining diagnostic accuracy and offers a scalable solution for the early detection and monitoring of neurodegenerative diseases in both advanced clinical and resource-constrained environments.

[0013] Furthermore, the combined scope of these inventions underscores a comprehensive approach to healthcare support, encompassing both physical assistance and diagnostic intelligence. While the assistive robotic system improves quality of life through real-time interaction, mobility support, and personalized care in a home environment, the diagnostic AI system provides physicians with precise, traceable, and efficient tools for the early detection of complex neurological disorders. Together, they represent a synergistic vision of the future of healthcare, in which intelligent machines not only support users in their daily lives but also contribute to timely and transparent medical decision-making, ultimately improving both patient autonomy and clinical outcomes.

Claims

[1] A multi-task deep learning system for the detection of Alzheimer's and Parkinson's disease using MRI, consisting of: a ResNet18-based Convolutional Neural Network (CNN) module configured to extract spatial features from preprocessed T1-weighted MRI scans; a first classification head coupled to the CNN module and configured to predict the disease type from a group consisting of Alzheimer's disease (AD), Parkinson's disease (PD) and a healthy control group; a second classification head which is conditionally activated by the first classification head when Alzheimer's disease is predicted and is configured to classify the Alzheimer's stage into one of the following stages: very mild, mild or moderate dementia; a combined loss function module configured to integrate categorical cross-entropy loss for disease classification and weighted cross-entropy loss for Alzheimer stage classification; an interpretability engine configured to provide predictive explanations and including the following: a Grad-CAM module for class-specific spatial activation mapping, a SHAP module for global feature mapping at the voxel level and a LIME module for the local, surrogate-based explanation of individual MRI scan predictions. [2] System according to claim 1, wherein the system further comprises a preprocessing module configured to scale MRI images to 224 × 224 pixels, normalize voxel intensities to a defined scale and standardize image orientation for consistent feature extraction. [3] System according to claim 1, wherein the interpretability engine comprises Grad-CAM for generating heatmaps, SHAP for determining meaning at voxel level and LIME for localized explanation using superpixel segmentation. [4] System according to claim 1, wherein the system comprises a premature termination module, a learning rate scheduler selected from StepLR or ReduceLROnPlateau and an Adam optimizer for efficient model training. [5] System according to one of the preceding claims, wherein the second classification head is only conditionally activated if the first classification head predicts Alzheimer's disease, thereby reducing the computational effort for non-Alzheimer's cases. [6] System according to any of the preceding claims, wherein the system is configured to be used in resource-constrained hardware environments, including edge devices, mobile diagnostic platforms or cloud-based inference servers, while maintaining real-time inference capability. [7] System according to any of the preceding claims, wherein the architecture is modular and containerized to be deployed using Docker and orchestrated using Kubernetes, thereby enabling integration into clinical workflows via cloud APIs. [8] System according to any of the preceding claims, wherein the classification results are exported together with interpretability outputs and confidence scores in a structured format that is compatible with electronic health record (EHR) systems via secure application programming interfaces (APIs).