Multispectral image fusion via heterogeneous integration of broadband image sensor and sub‑Boltzmann Transistors on CMOS Authors: Anupom Devnath, Batyrbek Alimkhanuly, Junseong Bae, Minwoo Lee, Jinsu Choi, Gisung Lee, Taemin Sim, Seungwoo Moon, Hyunwoo Sohn, Donguk Seo, Sandeep Kumar Maurya, Yoonmyung Lee, Seunghyun Lee Published: 2026 ABSTRACT AbstractBroadband multispectral imaging from ultraviolet to near-infrared enables spectral information fusion into a single composite, enriching inputs for machine vision and medical diagnostics. Conventional fusion systems require multiple photodetectors and co-registration algorithms, leading to resolution mismatch, area overhead, and computational burden, while readout circuitry increases chip area and power. Scaled transistors also suffer from off-state leakage and thermionic limits, increasing dark current and degrading signal-to-noise ratio. Here, we report a heterogeneously integrated quantum-dot-sensitized phototransistor active pixel sensor array incorporating series-connected CMOS-based steep-subthreshold switching transistors in a vertically stacked architecture. The switches operate below the Boltzmann limit, achieving a subthreshold swing of ~3.27 mV/dec over six decades, ~52-fold lower than CMOS, with ~0.1 pA leakage. The array exhibits ultralow dark current of ~36 fA, >120 dB linear dynamic range, and ~30 fJ/pixel switching energy. The phototransistor exhibits ~2 × 106 A/W responsivity and ~1.3 × 1014 Jones detectivity, enabling single-sensor pixel-level multispectral fusion for medical imaging.Subject terms: Electrical and electronic engineering, Optoelectronic devices and components, Electronic devices Introduction Recent advances in integrated circuit (IC) manufacturing1 have enabled high‑performance, energy‑efficient digital imaging systems with pixel densities now exceeding 200 megapixels per chip2. Monolithically integrating photonic components with CMOS chip technology promises even greater benefits, including broader communication bandwidths, higher connection densities, and compact on‑chip optical interconnects3,4. However, incorporating non‑silicon optoelectronic materials into silicon ICs remains a significant challenge, limiting the ability to fully leverage their potential for applications such as broadband multispectral imaging, on‑chip optical communication5, and highly area‑efficient image-sensing architectures4,6. Broadband multispectral imaging from ultraviolet (UV) to near-infrared (NIR) range has attracted increasing attention that enables pixel-level image fusion based on a single sensor platform, distilling the most informative content from different wavelength images into a single composite7. The result is richer input for downstream applications such as machine vision, autonomous systems, and medical diagnostics8. For example, in medical imaging, UV highlights surface texture, visible captures overall appearance, and NIR reveals deeper soft tissue9; thus, fusing them yields a more complete view for diagnosis. Unfortunately, conventional multispectral fusion systems utilize separate photodetectors per band placed at different pixel locations and then feed their outputs to vision algorithms (e.g., multiscale transforms10, deep learning11). This approach often introduces resolution mismatches, demands sophisticated computation, and increases system complexity8,12. Hence, a broadband multispectral CMOS-integrated active pixel sensor (APS) with on-chip readout, combining ultra-low dark current and high linear dynamic range, could enable high-fidelity pixel-level image fusion in demanding applications. A key step for a practical broadband image sensor array is the adoption of an active matrix with multiple transistors. Conventional CMOS APS designs, which use three (3 T) or four transistors (4 T) and a silicon photodiode13,14, continue to perform well for human-vision-range imaging on standard silicon platforms15,16. They offer moderate dark current and minimal image lag. However, as the diversity of broadband applications and the need for higher resolution increase, these architectures become less efficient in terms of cost per area, dynamic range, and energy efficiency. In response, 2T17–19 and even 1T20 APS concepts are gaining broad interest because of their compact footprint, simpler driving, and potential to extend imaging beyond visible light by applying alternative photodetectors that work on different wavelengths. Because a significant portion of the usable area in commercial APS devices is occupied by photo‑detecting silicon diodes, a substantial amount of the remaining chip area must be dedicated to readout circuitry2,14,15. During readout, the switching transistor transfers the signal to the column bus, where the column analog-to-digital converters (ADCs) digitize it. Notably, this data transfer and conversion process accounts for nearly 90% of the power consumption in conventional CMOS image sensors13, highlighting the need for circuit‑level improvements and more efficient use of the peripheral circuitry area. To improve system performance for fast-growing AI and IoT applications, the industry pursues both Moore’s law scaling and more than Moore strategies21–23—continued transistor miniaturization, and vertical heterogeneous integration/packaging on a single chip. Unfortunately, silicon miniaturization has reached a wall due to commonly well-known issues such as short-channel effects, process variations, and rising off-state leakage and static power1,21,24. In conventional transistors, source carriers follow a thermal (Boltzmann) distribution and cross a potential barrier, limiting gate control to the Boltzmann limit of about 60 mV per decade at room temperature, set by the thermal voltage (kT/q)25–27. High leakage in conventional APS also raises dark current even without light28, degrading contrast, and narrowing the dynamic range. Since shot noise scales with the square root of dark current, elevated dark current lowers the signal-to-noise ratio (SNR)29, especially under low light conditions, reducing sensitivity and detective quantum efficiency. If an emerging type of transistors25–27 with steep subthreshold swing and ultra-low leakage is to be applied, an innovative APS with fast switching at record low supply voltages, static power, and high linear dynamic range may become possible. Therefore, achieving sub‑Boltzmann operation through threshold switch assisted internal voltage amplification rather than fundamental thermionic violation of baseline transistor, together with ultra‑low subthreshold swing and leakage current may be a key step toward overcoming the limitations of conventional APS technology. In this work, we present a CMOS-integrated 32×32 array APS featuring a QD-coated indium-gallium-zinc-oxide (IGZO) transistor as a broadband phototransistor, paired with a steep-threshold switch in series with a 180 nm silicon CMOS FET that serves as the switching transistor. On the CMOS chip, the integration of ultra-thin, filamentary (atomic-scale) threshold switches (TS) with the underlying silicon FETs enables sub-Boltzmann transistors, characterized by ultralow subthreshold swing of 3.27 mV per decade and off-state leakage current of ~50 fA, supporting energy-efficient operation. IGZO was selected for its wide band gap, BEOL compatibility, and versatility30–32, offering ultra-low leakage and relatively high mobility, with clear advantages over polysilicon thin films and organic semiconductors31,33. For UV–NIR sensing, studies often report a sensitizing layer, with PbS quantum dots favored for their tunable band gap, size-dependent absorption, and broadband response16,34,35. The QD-coated IGZO phototransistor achieves broadband detection (300–1200 nm) across the UV-to-NIR spectrum, delivering responsivity of 2 × 106 A/W, detectivity of 1 × 1014 Jones, and a cut-off frequency of 106 kHz. At the pixel level, these features yield an ultra-low dark current (~36 fA), a noise floor of 1.89 × 10−15 A Hz−0.5, and a linear dynamic range exceeding 120 dB, outperforming typical silicon imagers (60–100 dB)13,36,37, while maintaining an energy consumption of only ~30 fJ per pixel. The common HfO2 insulator serves as the gate dielectric layer of the phototransistor and the electrolyte layer of TS, allowing facile monolithic integration. This integration further enables a streamlined 2 T pixel architecture in place of conventional APS designs, yielding a compact footprint of ~0.1024 cm² for 1024 pixels. Finally, as a proof of concept, broadband pixel-level multispectral image fusion is demonstrated, highlighting its potential for next-generation applications in medical diagnostics, machine vision, autonomous vehicles, and environmental monitoring. Results and discussion Integrated active pixel sensor design The fabricated CMOS chip, shown in Fig. 1a, is positioned on a fingertip and integrated onto a PCB carrier, allowing for efficient electrical interfacing and precise characterization in advanced optoelectronic applications. The functional architecture of the CMOS chip with a 32×32 active pixel array is depicted in Fig. 1b. A microscopic view of the lower right corner, accompanied by a scanning electron micrograph (SEM) of the active region (inset), reveals fine structural details. The right panel presents the standalone device, enabling direct, wirebond-free measurements for streamlined characterization. The standalone device enables independent characterization of either the switching transistor or the phototransistor, as well as single-pixel APS measurements, as shown in Fig. 1c. Figure 1d depicts the active array layout, exemplified here by a 5 × 5 pixel configuration for clarity. Figure 1e presents the single-pixel architecture, highlighting both the integration and end-of-line processes. The single pixel comprises a quantum-dot–sensitized IGZO phototransistor and a switching transistor; the latter is formed by the series connection of a threshold-switch (TS) element with a commercial 180 nm process silicon FET (TS-Si-FET). The switching transistor exhibits apparent sub-Boltzmann switching while suppressing leakage current, which is attributed to internal voltage amplification induced by the series integrated TS component rather than a fundamental violation of thermionic emission in the baseline Si-FET. To mitigate the persistent leakage currents in CMOS image sensors, even within established CMOS process technology, the TS-Si-FET architecture substantially improves the subthreshold performance leading to ultra-low power consumption, enabling near-negligible dark current in APS. Moreover, replacing the conventional photodiode with a phototransistor within the pixel architecture enables higher photoresponsivity, an improved signal-to-noise ratio, and reduced dark current. The TS configuration, along with its impact on electrical performance, is presented in the right panel. Figure 1f shows the SEM micrograph of the front-end-of-line (FEOL) transistor node on the manufactured CMOS chip. A schematic illustration of FEOL, BEOL, and the integration sequence is shown in Supplementary Fig. 3. Additionally, cross-sectional SEM images of the CMOS chip reveal the multilayer metal interconnects (metal-1 to metal-6 stack) above the 180 nm process. Figure 1g depicts the equivalent circuit of a single pixel in the active pixel image sensor array. The APS array performance for visible light is primarily shown in Fig. 1h. Figure 1i conceptually depicts the application of the configured APS for pixel-level multispectral image fusion, a process that extracts essential information from multiple wavelength sources into one unified image, enabled by the broadband detection of the QD-coated phototransistor, providing enhanced clarity and depth for applications ranging from medical diagnostics and biometrics to machine vision and artificial intelligence. CMOS-integrated broadband APS array (32 × 32) driven by steep slope switching transistors. a Photograph of the CMOS chip mounted on a PCB carrier (shown on a fingertip for scale). Scale bar: 7 mm. b Optical micrograph of the APS highlighting functional regions; the image was taken before integrating the TS cell and phototransistor to improve contrast. Insets show the optical micrograph of array region with SEM result. Scale bars: 500 µm, 200 µm, and 200 µm, respectively. c standalone on-chip test structures used to electrically characterize the QD-coated IGZO phototransistor, the TS cell, and the 180 nm baseline Si-FET in the end-of-the-line process. Scale bar: 50 µm. d Layout of the active area and readout circuitry, including the array measurement node and pixel configuration (partial array view). e Top view schematic of a single pixel APS illustrating BEOL integration and via-connected metal routing; each pixel integrates a broadband phototransistor for sensing and a baseline Si-FET in series with a TS element to form a steep-slope switching transistor. The integration and end-of-the-line process are marked. TS cell structure and conceptual functionality in switching transistor over conventional transistor are shown in inset. f SEM image of the 180 nm process silicon transistor (Si-FET) implemented in the CMOS chip. Scale bar: 200 nm. g Single pixel circuit schematic. h Representative image sensing results acquired with the CMOS APS array. i Implementation schematic for broadband, energy-efficient, fast-switching CMOS APS operation for single-sensor based pixel-level multispectral fusion in medical imaging. Performance of ultralow power switching transistors CMOS image sensor design is increasingly constrained by leakage currents, with scaling to 180 nm exacerbating off-currents and lowering supply voltages14,38,39. These trends, coupled with the difficulty of further voltage scaling, intensify power dissipation and degrade the on/off current ratio. An efficient way to minimize power consumption is to achieve a steep subthreshold swing (SS) with reduced leakage current at low supply voltage (further details are provided in Supplementary Note 1)25,27. A threshold-switching silicon FET (TS-Si-FET) with sharp on/off characteristics and ultra-low energy consumption is employed as the switching transistor, enabling an energy-efficient APS with enhanced optoelectronic performance. Figure 2a shows the circuit diagram of the switching transistor (TS-Si-FET), in which a baseline silicon FET is integrated in series with a TS element. Figure 2b presents the high-resolution transmission electron microscopy (HR-TEM) cross-sectional view of the TS element, accompanied by Energy-dispersive X-ray spectroscopy (EDS) elemental mapping and the corresponding line profile. The Ag layer, HfO2 dielectric, and top and bottom electrodes (TE/BE) are distinctly visible, with the Ag atomic layer localized at the TE/HfO2 interface. Figure 2c and d present the typical I−V behaviors and cumulative voltage distributions of the TS device, measured under a compliance current (ICC) of 100 μA. The device exhibits volatile threshold-switching behavior, characterized by a low threshold voltage (Vth−TS ~ 0.26 V, Vth−TS is the threshold voltage of the TS), ultralow leakage current (<0.1 pA), and an on/off current ratio exceeding 109. It transitions to the on-state when the applied voltage surpasses Vth−TS and reverts to the off-state when the applied voltage drops below the hold voltage (Vhold−TS is the hold voltage of the TS). The switching operation of the TS device and the corresponding filament morphology are shown in Fig. 2e. The switching slope is remarkably steep at <1 mV/dec, with cyclic measurements showing Vth−TS consistently in the range of 0.176-0.219 V. Applying a sufficient voltage to the TS triggers abrupt Ag filament formation, which typically induces a negative differential resistance (NDR) effect across the device. The I−V characteristics measured under ramped current-sweeping mode, shown in Fig. 2f, confirm robust current-controlled (S-type) NDR behavior40–42, indicating a volatile and abrupt formation process. The volatile, rapid resistance transition of the atomic Ag filamentary TS generates an NDR effect that internally amplifies the voltage drop across the silicon FET channel, thereby driving a steep-slope conductance transition and enabling the device to surpass the fundamental thermionic limit. The swift formation and self-rupture of the atomic Ag filament within the TS enable the TS-Si-FET to exhibit sharp on/off switching and an ultralow SS. Electrical characterization of the series threshold switch integrated silicon switching transistor (TS-Si-FET) in the CMOS APS. a Single-pixel circuit; inset, TS-Si-FET formed by a baseline Si-FET in series with a threshold-switch (TS) element. VG is the Si-FET gate (G) − source (S) bias and VD is the drain (D) − source bias across the series pair; the internal node D* defines the voltage drop across the baseline Si-FET. b TS stack microstructure: HR-TEM (left) and cross-sectional EDS mapping with a line profile resolving Pd, Ag, Hf, and Cr (right). Scale bar: 10 nm. c Forming free I–V characteristics of the TS over 50 cycles (100 μA compliance), yielding ultralow leakage (<0.1 pA). d Cumulative distributions of Vth−TS, Vhold−TS and ΔV(Vth−TS−Vhold−TS), indicating reliable switching performance. e Schematic of voltage-driven Ag filament formation and rupture within the TS device. f Operating mechanism of the TS component. TS cell in series with a 100 kΩ load resistor (RL) exhibiting snapback switching. During a current sweep, the device traverses three states: (1) an insulating TS medium, (2) initiation and growth of an Ag filament, and (3) conduction sustained by the formed filament. g Band diagrams of the TS-Si-FET at thermal equilibrium for VGVth(right). Electrons traverse the contact barriers by thermionic emission or tunneling and cross the electrolyte barrier by hopping. h Forward-sweep transfer characteristics (ID−VG) of the baseline Si-FET and TS-Si-FET measured with a voltage step size of 10 mV over a sweep range from −0.75 V to 1.25 V. i Extracted SS in forward sweeps for the Si-FET and TS-Si-FET. The baseline FET remains above the thermionic limit, whereas the TS-Si-FET achieves a minimum SS of 3.27 mV per decade across >6 decades of drain current. j Off-state leakage and SS comparison highlighting abrupt switching and low-power operation. Violin plots show data distribution; boxes indicate 25th–75th percentiles, whiskers show 1.5 × interquartile range, center lines show median, stars show mean. k Output characteristics (ID−VD) of both devices, showing distinct switching behavior. l Output characteristics of TS-Si-FET at different compliance current. In the TS-Si-FET, the drain-to-source voltage (VD) is applied across both the TS and the Silicon FET channel, while the gate-to-source voltage (VG) governs the switching behavior. The applied VD is distributed between the TS and the FET, with VD representing the total device voltage drop and VD* the voltage drops across the Silicon FET. In this series configuration, VG adjusts the Fermi level of the Silicon FET channel, providing precise control over channel resistance (Rch), which largely determines the voltage drop between the FET and TS. The TS-Si-FET band diagram at a representative gate voltage relative to its threshold voltage (Vth) is depicted in Fig. 2f. The HfO2 electrolyte layer in the series TS operates as a VG-dependent barrier to electron flow, governing the voltage distribution of the channel and the TS. At VG < Vth, the channel Fermi level undergoes only a slight downward shift, maintaining high Rch. In this state, the TS current is insufficient to initiate Ag filament formation, with the HfO2 electrolyte imposing a strong barrier to electron transport. Once VG exceeds Vth, the Fermi level drops significantly, driving the channel into a low-resistance state and allowing TS current to trigger Ag filament formation, resulting in an abrupt rise in current. The transfer characteristics of the baseline Si-FET and the TS-Si-FET are presented in Fig. 2g. The baseline Si-FET exhibits an on/off current ratio of 3 × 107 with an SS of 105 mV/dec, exceeding the thermionic limit of 60 mV/dec at room temperature (Fig. 2h). Electrical characterizations of the baseline Si-FET are shown in Supplementary Fig. 4. In contrast, the TS-Si-FET demonstrates superior characteristics, with a higher on/off ratio (>5 × 1010), an off-state (or leakage) current as low as 5 × 10−14 A, and an ultra-steep SS of 3.27 mV/dec (maintained across the current range from 6.95 × 10−14 to 4.05 × 10−8). The key lies in the series integration of a TS, enabling internal amplification that overcomes the Boltzmann limit. The NDR effect of the TS cell from a physical standpoint is provided in Supplementary Note 2. The near-ideal switching (SS ~ 0 mV/dec) performance of TS-Si-FET is demonstrated in Supplementary Note 3. Importantly, the ultrahigh resistance of the TS dominates the off-state behavior, reducing ID to an ultralow level of 5 × 10−14 A. The extracted SS for both baseline and TS integrated Si-FET are shown in Fig. 2i. Figure 2j reveals that the TS-Si-FET markedly outperforms the baseline Si-FET device, reducing the SS by 52 times (right) and lowering the off-state leakage current by roughly 400 times (left). The device-to-device transfer characteristics with both forward and reverse sweep of the switching transistor are shown in Supplementary Fig. 5, with five repeated cycles measured for each device. Device-to-device and cycle-to-cycle variation, along with SS distributions for switching transistors, are presented in Supplementary Fig. 6. Moreover, the pixel-to-pixel SS distribution measured from 64 randomly selected standalone APS pixels is shown in Supplementary Fig. 7 and Supplementary Fig. 8. Figure 2k shows a clear distinction between on- and off-states, with conductance dominated by the TS for VD < Vth−TS, by the channel for VD ≥ Vth−TS, and exhibiting a sharp drop once VD falls below Vhold−TS. The output characteristics of TS-Si-FET for different compliance currents are shown in Fig. 2l. In-depth circuit-level analysis of the TS-Si-FET switching process in both forward and reverse sweeps is shown in Supplementary Fig. 9. Due to the volatile switching nature of the TS cell, the switching transistor exhibits low hysteresis, which can be reduced to a negligible level under higher drain bias (Supplementary Fig. 10). Reliability and durability tests, such as time-dependent DC pulse stressing and DC pulse tests of the switching transistor below and above the threshold voltage, are shown in Supplementary Fig. 11. Moreover, the temperature dependent performance of the switching transistor was evaluated from 298 to 348 K, as shown in Supplementary Fig. 12, confirming its nearly stable operation under thermal variation. Performance benchmarks of the different steep-slope transistor devices, in terms of SS versus on–off ratio and SS versus applied voltage, are illustrated in Supplementary Fig. 13 and show positioning in a promising zone for highly efficient, low-power circuitry. Phototransistors performance Phototransistor channel material (IGZO) characterizations with high-resolution XPS and elemental mapping are presented in Supplementary Fig. 14. As noted above, quantum dot sensitization is a well-known approach for broadband photodetection, and Supplementary Fig. 15 presents PbS QD characterization, including planar TEM, absorption spectra, and XPS survey spectra. Pixel-to-pixel variation in the photoresponse of the APS QD-sensitized phototransistor was examined, with Fig. 3a showing the transfer characteristics (ID−VG) of phototransistor devices measured before and after illumination at 520 nm (green) with an intensity of 1 mW cm−2. Figure 3b–e presents histograms of the distributions of dark current (Idark), photocurrent (Iphoto), photocurrent-to-dark current ratio (rPH=Iphoto/Idark), responsivity (R=Iphoto/WLPin; W, L, Pin denotes the channel width, channel length, and laser power, respectively), and detectivity (D*=A/NEP;A,andNEP denotes the area of the phototransistor, noise equivalent power by NEP=Inoise/R, respectively and calculated at 1 Hz) at VG = −0.5 V. The mean values were ~2.31 × 10−13 A µm−1 (Idark), ~ 3.68 × 10−8 A µm⁻¹ (Iphoto), ~ 4.90 × 104 (rPH), ~ 3.69 × 103 A W−1 (R), and ~1.16 × 1011 Jones (D*), with standard deviations of ~1.72 × 10−13 A µm−1, ~2.81 × 10−8 A µm−1, ~4.75 × 104, ~2.04 × 103 A W−1, and ~6.76 × 1010 Jones, respectively. Upon illumination, PbS quantum dots generate electron–hole pairs. Under positive gate bias (VG > 0), as depicted in Fig. 3f, downward band bending in the IGZO channel lowers the QD/IGZO interfacial barrier and promotes rapid injection of photoelectrons into the IGZO conduction band, while the large valence-band offset confines holes in the QDs. These long-lived holes act as a photogate, capacitively modulating the channel and yielding a high drain current. In contrast, a negative gate bias (VG < 0) depletes the IGZO channel, raises the interfacial barrier, and suppresses electron transfer across the heterojunction, resulting in a markedly reduced current. The photo-responsive transfer characteristics for different wavelengths across the visible spectrum are shown in Fig. 3g. The extracted responsivity under different light intensities is shown in Fig. 3h. The linear dynamic range was defined according to standard CMOS imager convention as LDR = 20logPmax/Pmin, where Pmax and Pmin are the maximum and minimum incident optical power densities within the measured linear response range. It quantifies the span of scene luminance over which a phototransistor’s photocurrent scales linearly with incident intensity and thus sets the range of dark-to-bright tones that can be reliably captured. Figure 3i shows a linear increase in photocurrent with increasing intensity from UV to NIR, yielding an LDR exceeding 120 dB for visible light at pixel level measurement, outperforming typical silicon imagers (60–90 dB)13,36,37. Under 520 nm illumination, the phototransistor exhibits fast transients (Fig. 3j) with rise (τrise) and fall times (τfall) of 1.36 and 1.42 μs, respectively. The rise and fall time were extracted using standard definitions43, with τrise=τ90%−τ10% and τfall=τ10%−τ90%. The normalized transient response of phototransistor over 2000 s to confirm its strong retention stability, as depicted in Supplementary Fig. 16. The inset frequency response shows a − 3 dB cutoff (f−3dB) above 100 kHz, consistent with the time-domain measurements. Benchmarking the figures of merit of the QD-sensitized IGZO phototransistor against those of reported devices across diverse materials and architectures reveals that our device achieves ultrahigh detectivity while maintaining responsivity comparable to the state of the art, as shown in Fig. 3k. The extracted data are shown in Supplementary Table 1. Optoelectrical characteristics of QD-coated IGZO phototransistor with device-to-device variation. a Transfer characteristics of IGZO phototransistors under pre-illumination and post-illumination. Source to drain current (ID) as a function of back gate voltage (VG) measured at a source to drain bias (VD) of 0.25 V. Transfer curves (ID−VG) of IGZO phototransistors measured before and after 520 nm illumination (1 mW cm⁻2). Extracted device-to-device variation of phototransistor key parameters. Histograms of dark current and photocurrent (b), Photo to dark current ratio before and after illumination (c), Responsivity (d), and detectivity (e). f Operating principle of the phototransistor under illumination at negative and positive gate bias. g Photo-induced transfer curves plotted on linear and logarithmic scales for blue (405 nm), green (520 nm), and red (638 nm) illumination across varying intensities. h Responsivity as a function of incident light intensity under 520 nm illumination. i The usable range from dark to bright, quantified by the linear dynamic range (LDR) at pixel level, exceeds 120 dB as the incident intensity increases from ~1×10⁻9 to 1×10⁻3 W cm⁻2, and it grows further as the measurement range extends to higher intensities. The inset shows the LDR at different wavelengths. j The transient response was characterized using the rise and fall times. Left inset: measured pulsed response. Right inset: cut-off frequency, which is consistent with time-domain measurements. k Literature benchmark of responsivity and detectivity demonstrates that our device achieves ultrahigh performance relative to prior devices fabricated with diverse materials and structures. Active pixel image sensor optoelectronic performance The simplified 32 × 32 CMOS integrated array and its single pixel equivalent circuit are shown in Fig. 4a. We evaluate optical switching by pulsing Vsw from 0 to 0.3 V at fixed supply and phototransistor-gate biases while sequentially illuminating the pixel with equal-power light across the visible spectrum (1 mW cm−2, 5 s per wavelength). The response reveals strong switching fidelity and a wavelength-tuned current level that decreases from blue to red, as depicted in Fig. 4b. With a clear separation between the pre-switching under dark and post-switching under light conditions, Fig. 4c presents the current output signals for all 1024-pixel units, suggesting that crosstalk is minimal. The inset dark current histogram (mean ~36 fA; standard deviation ~18 fA) underscores an ultralow dark current, attributable to the suppressed leakage of the TS integrated switch. Before switching, the TS element remains in its high resistance state and acts as the dominant current limiting component in the pixel. As a result, the dark current is mainly suppressed by the TS element in the pre switching regime. The output current of the 32×32 APS under pre-switching dark and post-switching light conditions, with corresponding dark-current and photocurrent distributions, is mapped in Supplementary Fig. 17. The dark current of the proposed APS is benchmark-setting, outperforming the ultra-low dark current regime pursued in leading photodetectors. Additionally, Fig. 4d presents the current output of the single pixel switching response under broadband illumination from UV to NIR. Owing to the ultra-low dark current of the APS, the noise current (Inoise) is 1/ƒ dominated at low frequency (~4.13 × 10−15 A Hz−0.5), as depicted in Fig. 4e, and higher than the shot noise limit at higher frequencies (~3.4 × 10−16 A Hz−0.5). The thermal noise current density, estimated as ithermal=4kBT/R, is ~4.5 × 10−17 A Hz−0.5 at room temperature, which is lower than the both shot noise and the measured pixel noise. Therefore, the measured noise is likely dominated by trap/interface related 1/ƒ noise and readout or instrumentation noise, rather than by thermal or shot noise alone. A key metric of the active pixel sensor is its energy consumption, which can be evaluated using the following equation. Pixel- and array-level sensing metrics for a CMOS-integrated 2 T APS. a Simplified CMOS integrated array with the single pixel APS equivalent circuit. b Photoinduced switching characteristics of 2 T APS under the switching voltage and visible (RGB) light pulses. c All-pixel output currents acquired before and after switching, under both pre- and post-illumination conditions. d Output current for single pixel switching under broadband illumination from UV to NIR. Data are presented as mean ± standard deviation. e Noise current single pixel APS as a function of frequency. f Distribution of the calculated per-pixel readout energy consumption for the proposed APS and a control APS in which the switching transistor omits TS integration. Data points show individual measurements, center lines show means, and whiskers show ranges. g Application to optical communications. Top panel: schematic diagram of the UV and NIR optical communication system. Bottom panel: transmission of ASCII code information using a single APS as the signal receiver under UV and NIR light illumination. h Schematic illustration of the image capturing using a CMOS APS array. Photo-to-dark current ratio mapping from image detection under RGB illumination reveals the butterfly image. Here, ID denotes the current (from ground to the supply) drawn by the APS under illumination. VG1 and VG2 are the gate biases of the switching transistor and the phototransistor, respectively. The corresponding gate capacitances, CG1 and CG2, follow CG=ε0εoxWL/tox, where ε0 is the vacuum permittivity, εox the dielectric relative permittivity, tox the gate-dielectric thickness, and W and L the channel width and length of the respective devices. The first two terms of Eq. (1) correspond to charging the gate of both transistors, and the third term reflects the energy lost in the phototransistor from the total ID (total of the dark and photocurrent) flowing during the illumination time (texp). The pixel-level readout energy (Ereadout=VD∫0tpulseID(t)dt) is calculated from the measured current during pulsed APS operation. To examine how the TS affects per-pixel energy, we fabricated a control APS using a switching transistor without a TS and an unchanged phototransistor. Figure 4f highlights per-pixel readout energy consumption for the proposed APS versus a control, indicating a roughly twenty-fold improvement in energy efficiency due to suppression of dark current by the TS. Supplementary Table 2 benchmarks the proposed CMOS APS with prior APSs (excluding commercialized conventional silicon photodiode-based structures) and shows substantial improvements in integration scheme, detection range, energy per pixel, dark current, and dynamic range. To demonstrate optical communication, a host computer encoded the target string KHU in ASCII code, forwarded the bit stream to a signal generator, and amplitude-modulated lasers at 365 nm (UV) and 850 nm (NIR). The modulated beams were directed onto the device, which acted as the receiver; the pixel output generated electrical pulses that were streamed back to the computer for decoding (acquisition trace in Fig. 4g, top). After processing, the ASCII sequence was correctly reconstructed as KHU (Fig. 4g, bottom). The square-wave photocurrent closely reproduces the transmitted bit pattern, confirming the device’s suitability for ultraviolet and infrared optical communication. To investigate the image-capturing properties of the CMOS APS array, we used RGB illumination with corresponding shadow masks, with the top, middle, and bottom masks applied to green, blue, and red, respectively, to selectively block light entry (the turn-on process) and implement the designed framework, as schematically depicted in Fig. 4h. This configuration produced the intended butterfly image (Fig. 4h) at a resolution of 1,024 pixels. Note, since the silicon-based switching transistor is beneath the phototransistor, its operation remains free of light interference during imaging. Implementation of multispectral image fusion Multispectral image fusion integrates images acquired at different wavelengths, distilling the most informative content into a single composite image that provides a stronger basis for subsequent analyses and applications. Medical image fusion combines images from a single modality or across modalities (e.g., X-ray, MRI, CT, PET; Fig. 5a) to improve diagnosis and clinical assessment, and interest in these methods continues to grow12,44. Conventional multi-sensor-based image fusion often requires separate imaging paths, image registration and post processing and often leads to pixel mismatch and high computational load. In contrast, an ultrafast, energy-efficient CMOS APS array that performs pixel-level multispectral fusion across ultraviolet, visible, and infrared bands could enable co registered spectral information within single sensing platform and support biomedical analysis and diagnosis. A schematic of a multispectral pixel-level image fusion mechanism based on a single-sensor platform is shown in Fig. 5b. The transmission mode imaging process shows UV, visible, and infrared sources, a fabricated finger mimetic device as an object, and a CMOS APS array as the detector. Since the CMOS chip is ultra-small (7 mm × 7 mm), we mimicked the finger structure in several layers (fingerprint ridges, dermis, and veins) using ZnO for UV absorbance, perovskite QD for visible light absorbance, and PbSe QD for NIR absorbance, as shown in Fig. 5c. The absorbing layers were fabricated on a transparent glass substrate, and their planar areas were scaled in descending order (ZnO