Detection of Multiple Sclerosis Based on MRI Images using a Low-Power Analog Integrated Artificial Neural Networks

Artificial Neural Networks (ANNs) have become increasingly prevalent for classification and decision-making tasks across various engineering domains. However, their high computational resource requirements remain a significant barrier, particularly for portable or embedded applications. To address this challenge, this paper proposes an innovative method where the ANN learning phase is conducted using conventional software-based training, while the classification phase is implemented as an analog electrical circuit derived from the trained model. The method is validated on a medical diagnostic application: the detection of Multiple Sclerosis (MS) lesions from magnetic-resonance imaging (MRI) images. The dataset comprises 2,831 FLAIR and T2-weighted MRI images, partitioned into 70% training and 30% testing subsets. The proposed analog circuits achieve accuracies of 93.31% (MLP) and 94.82% (ResNet-10), while significantly reducing power consumption to 728 nW (MLP) and 967 nW (ResNet-10) and inference latency to 0.2 µs. Although the achieved accuracy slightly trails that of purely software-based implementations, the substantial reductions in power usage and response time make this hardware-based approach highly advantageous for real-time, low-power medical diagnostic devices.