SCIENCE CHINA Materials•2026•DOI: 10.1007/s40843-025-3786-8
Conventional cancer diagnostic techniques, such as tissue sampling and microscopy, are invasive and prone to misdiagnosis, driving the need for non-invasive, precise alternatives. Chiral biophotonics, exploiting circularly polarized light (CPL), offers unique polarization-selective interactions with biological tissues, enabling higher imaging contrast and molecular-level discrimination. However, current CPL detection technologies are passive and single-mode, lacking dynamic tunability and parallel processing capabilities. Meanwhile, AI-assisted diagnostics rely on separated sensing and computing units, suffering from poor integration and transmission inefficiency. Here, we report a near-infrared (NIR) chiral organic synaptic photodiode with electrically tunable dual-mode operation, enabling simultaneous CPL detection and neuromorphic processing. Under negative bias, the device operates as a highly sensitive CPL detector for chiroptical signal acquisition. Under positive bias, it exhibits history-dependent synaptic behavior with photocurrent dissymmetry factor (g_ph) dynamically tunable up to -0.06. By integrating this device into an optical convolutional neural network (OCNN), we achieved intelligent cancer detection with CPL-based imaging. Experimental results demonstrate that CPL detection accuracy reaches 83%, approaching the theoretical 87%, significantly outperforming natural light detection at 65%. The device enhances image contrast and feature extraction, laying a foundation for intelligent, adaptive diagnostic systems.
SCIENCE CHINA Materials•2025•DOI: 10.1007/s40843-025-3576-6
Emulating biological synaptic plasticity is essential for advancing artificial intelligence. However, in most existing synaptic phototransistors to date, both electrical and optical stimuli induce weight modulation within a comparable dynamic range, limiting plasticity tunability and richness. Here, we report a synaptic phototransistor that enables distinct weight modulation in response to electrical and optical inputs, achieving hierarchical, multi-scale plasticity with concurrent visible-light emission for direct display. The device integrates a long-afterglow material that converts transient ultraviolet (UV) excitation into persistent visible emission, serving as a temporally extended, memory-like optical stimulus. Compared to direct electrical gating, this delayed optical activation of the optoelectronic channel induces weight modulation on a significantly longer timescale, enabling hierarchical plasticity and cascade interactions between optical and electrical pathways. The dual-output architecture allows simultaneous optical visualization and electrical signal processing, effectively integrating optical perception with in-sensor computation. Leveraging this design, we demonstrate a UV-resolvable neural network capable of direct image display and achieving a recognition accuracy of 95.03% for handwritten digits. This work establishes a new paradigm for multimodal neuromorphic systems by seamlessly integrating sensing, display, and computation within a unified in-sensor architecture.
SCIENCE CHINA Materials•2025•DOI: 10.1007/s40843-025-3524-y
The 12-lead electrocardiogram (ECG) is indispensable for the initial diagnosis of cardiac conditions, yet existing neuromorphic hardware for multi-lead ECG monitoring requires multiple array circuits and two operational processes, imposing severe constraints on device consistency and diagnostic accuracy. This study introduces a neuromorphic parallel computing hardware architecture based on quantum dot synaptic transistors that leverages trap and surface electric field effects to enable 12-lead ECG monitoring within a single array circuit, eliminating the need for twelve separate circuits. The system concurrently processes multiple ECG signals and produces final outputs without external computing or control circuits. A 12-transistor array, termed STAC, directly processes one-dimensional ECG data without additional conversion circuits, integrating a feature extraction layer at the pixel level and a feature fusion layer at the circuit level. Classification of ECG signals from the MIT-BIH Arrhythmia Database and the Chinese Twelve-Lead ECG Challenge Database yields a training accuracy exceeding 98%. A five-class ECG signal classification task achieves 96.2% recognition accuracy, with a 5×5 confusion matrix confirming high classification precision across normal (N) and four abnormal categories (A, V, L, R). The architecture accurately detects myocardial infarction by fine-tuning internal weights, demonstrating proficiency in monitoring abnormal ECG signals. This advancement offers a compact, low-cost solution for wearable and portable 12-lead ECG monitoring devices, enabling real-time cardiac assessment with reduced hardware complexity and enhanced diagnostic reliability.