Energy-efficient high-fidelity image reconstruction with memristor arrays for medical diagnosis

Abstract

Medical imaging is an important tool to make accurate medical diagnosis and disease intervention. Current medical image reconstruction algorithms mainly run on Si-based digital processors with von Neumann architecture, which faces critical challenges to process massive amount of data for high-speed and high-quality imaging. Here, we present a memristive image reconstructor (MIR) to greatly accelerate image reconstruction with discrete Fourier transformation (DFT) by computing-in-memory (CIM) with memristor. To implement DFT on memristor arrays efficiently, we proposed a high-accuracy quasi-analogue mapping (QAM) method and generic complex matrix transfer (CMT) scheme, to improve the mapping precision and transfer efficiency, respectively. With these two strategies, we used MIR to demonstrate high-fidelity magnetic resonance imaging (MRI) and computed tomography (CT) image reconstructions, achieving software-equivalent qualities with peak signal-to-noise ratios (PSNR) of 40.88 dB and 22.38 dB, respectively. The reconstructed images were then segmented using a popular nnU-Net algorithm to further evaluate the reconstruction quality. For the MRI task, the final DICE scores were 0.979 and 0.980 for MIR and software, respectively; while for the CT task, the DICE scores were 0.977 and 0.985 for MIR and software, respectively. These results validated the feasibility of using memristor-reconstructed images for medical diagnosis. Furthermore, our MIR also exhibited more than 153× and 79× improvements in energy efficiency and normalized image reconstruction speed, respectively, compared to graphics processing unit (GPU). This work demonstrates MIR as a promising platform for high-fidelity image reconstruction for future medical diagnosis, and also largely extends the application of memristor-based CIM beyond artificial neural networks.

Linear, symmetric, self-selecting 14-bit molecular memristors

“…The quest for an efficient general-purpose VMM engine has led to hardware advances, notably the dot-product engine (DPE) -an analogue accelerator that enables VMM in a single time-step [17][18][19] . Despite its transformative potential across computing scales, the adoption of DPEs has been limited by their low precision, stemming from analogue circuit elements offering only 2-6 equivalent bits 10,11,[17][18][19] . This lack of precision, rooted in physical non-idealities including nonlinear weight updates, asymmetric behaviour, noise, conductance drifts, and device-to-device variability, is a foundational challenge in neuromorphic computing 20,21 .…”

Nonlinear Memristive Computational Spectrometer

“…In this work, photonic memristors [38][39][40][41][42][43] emerge as a groundbreaking alternative, utilizing their distinctive energy band coordination and memory effects for the high-performance spectrometer design. By leveraging the nonlinear properties of memristors, we introduce a PIN WSe 2 homojunction-based nonlinear spectrometer that offers high performance and resolution within an ultra-compact footprint.…”

A habituation memristor model for lung cancer screening application

“…Since habituation allows for filtering out redundant information and thereby emphasize the important one, it would offer an effective mechanism for many cognitive tasks such as disease or cancer findings [18][19][20]. The non-linear characteristics and programmability of memristor circuits have enabled significant advancements in the implementation of artificial intelligence, medical diagnosis and screening [21][22][23]. For example, a real-time and in situ medical diagnosis method using memristor circuit cores has been proposed [22].…”