Publication

Abstract : Alzheimer is one of the key mental illnesses suffered by 55 million people worldwide and 10 million new patients are added every year. If Alzheimer’s is detected in early stage, then the impact of disease will be reduced. Artificial intelligence (AI) models are helping in early diagnosis using MRI images. For better AI model development, the model needs to be trained over a huge dataset. But Health data privacy is another key concern of data collection, and hence federated learning framework helps models get trained over data without data transferring to cloud platform. Other Key concerns of AI model development is the availability of labeled data and benchmarking the model performance developed on dataset, which is trimmed or is a modified version. As availability of research work on specific data is rare. Currently, researchers have to develop other research models for specific data and such approaches take more time. Also, doctors expect explainability of the model, instead of black box model. Proposed BrainCrossFed model addresses data privacy concerns by avoiding the need for data transfer to cloud and leverage available labeled data at different federated nodes along with a dataset for Alzheimer to compute cross fed average of model weights. Due to the proposed Cross Fed Avg algorithm performance of the proposed deep learning model accuracy is increased from 99.11% to 99.77% close to 100%. Also such cross fed average enables the model to find global minima. Final performance achieved by the proposed model is 99.77%, with 100%. The performance of the model is benchmarked with current well known classification models, State-of-the-Art (SOTA), self supervised models like DINOv2 [1], for the same datasets. DINOv2 model has achieved 98%. Other generic well known models like ResNet101, DenseNet121, and Visual Geometry Group16 (VGG16) and customer models [2] have achieved max performance 97.60%. The GradCAM developed to explain the classification result.