A Cross-Scale Memristor-Based Neuromorphic Platform Bridging Physical Encoding and Medical Inference

Despite the progress in machine vision, existing image recognition systems still operate in isolation from the physical world, lacking dynamic temporal encoding, in-memory processing, and intrinsic interpretability. Here, a memristive vision architecture that integrates multimodal sensing, physical memory, and medical inference within a unified device-level framework is developed. Specifically, spatial gradients from retinal fundus images are encoded into Ag/Indigo/FTO memristor array via pulse modulation, generating conductance maps (“G maps”) that serve both as nonvolatile memory and as neuromorphic inputs for lesion segmentation via lightweight U-Net models. Beyond static vision, the same memristor exhibits strong temporal sensitivity, enabling biologically realistic nociceptive encoding and Morse code recognition. These results demonstrate a cross-task, cross-scale system that merges physical encoding with machine intelligence, bridging the gap between material dynamics and high-level perception. Therefore, this paradigm offers a scalable, interpretable hardware foundation for next-generation bio-perceptive computing and intelligent diagnostics.