Neural compressed sensing for MRI and ECG data acquisition and compression

Biomedical signal acquisition often occurs in resource-constrained environments, necessitating advanced encoding or acquisition algorithms. In such contexts, Compressed Sensing (CS) offers a promising solution but faces performance challenges, especially in practical implementations. This dissertation explores the integration of Neural Network and Compressed Sensing techniques for the efficient acquisition and compression of biomedical signals, specifically focusing on Electrocardiogram (ECG) and Magnetic Resonance Imaging (MRI) data. For ECG, two innovative approaches are presented. The first approach introduces a data-driven binary encoding to develop a lightweight encoding mechanism. The second approach introduces an adaptive, incremental compression scheme that uses a performance predictor to dynamically adjust the number of transmitted measurements. For MRI data acquisition, the work delves into advanced undersampling techniques. The first part builds on the state-of-the-art LOUPE architecture, incorporating CS-derived constraints into the training framework to improve the quality of reconstructed MRI images. The second part, introduces the concept of incremental acquisition, where the number of acquired k-space samples is dynamically adjusted based on real-time quality assessments. This dissertation demonstrates how combining model-based CS with data-driven DNN holds the potential to revolutionize acquisition methodologies for biomedical signals, making advanced diagnostics efficient even in resource-hungry settings.