Shenzhen University team builds photonic AI system for faster medical diagnosis

Researchers at Shenzhen University and industry partners have built an all-fiber photonic AI platform that uses light instead of electrons to diagnose medical images faster and with far lower energy use. In tests on retinal detachment and liver cancer, the system matched senior radiologists while cutting latency and power demand. Why it matters: - The system points to a lower-energy path for AI in healthcare, where current GPU-based tools can be slow, hot, and power-intensive. - Faster image analysis could improve diagnosis in time-sensitive settings such as rural clinics, ambulances, and other resource-limited sites. - The work also suggests a route to greener AI by reducing the carbon footprint of medical computing. What happened: - Shenzhen University researchers, working with Shenzhen Metasensing Technology Co., Ltd. and Shenzhen All-Optical Era Technology Co., Ltd., developed a black phosphorus-based all-fiber photonic AI diagnostic platform. - The project was led by Professor Han Zhang from Shenzhen University’s College of Physics and Optoelectronic Engineering. - The results were published online on May 28, 2026, in Opto-Electronic Advances under the title, “Tunable phosphorene modulators – accelerating medical diagnosis with ultra-efficient photonic platforms.” - The co-first authors were Dr. Yi Liu, Huide Wang, and Honghai Zhu. - The corresponding author was Prof. Han Zhang. The details: - The platform uses a microfiber knot resonator, a black phosphorus and molybdenum disulfide heterostructure, and a Ring-Assisted Mach–Zehnder Interferometer architecture. - The black phosphorus and MoS2 layers were assembled as a van der Waals heterostructure and integrated onto an optical fiber structure thinner than a human hair. - The microfiber knot resonator increases light-matter interaction, allowing a tiny voltage to shift the light wavelength and enable optical modulation. - The team fabricated the heterostructure by mechanical exfoliation and dry transfer, then integrated it onto a microfiber knot resonator hundreds of micrometers wide and tapered to a few micrometers at the waist. - Two RAMZI units plus a photoreceiver were combined in a time-division multiplexing setup to create a full all-fiber photonic neural network. - In the HCC task, the researchers used 3,348 dynamic contrast-enhanced CT studies, including 2,458 biopsy-confirmed liver cancer cases and 890 normal controls. - The system reached 95.0% accuracy and 97.6% specificity on hepatocellular carcinoma diagnosis. - In the retinal detachment task, the team used 40 desensitized B-scan images of retinal detachment and 40 images of normal retinas. - The system matched experienced radiologists in the clinical validation tasks. - Processing one liver CT study took 85 milliseconds on an NVIDIA A100 GPU and 0.8 milliseconds on the photonic system. - Energy use was 0.608 femtojoules per operation on the photonic system versus 150 femtojoules on the NVIDIA A100 GPU, a 246-fold gain in energy efficiency. - The system’s optical fiber transmission loss was below 0.2 dB/km. - The RAMZI device achieved a modulation efficiency of 0.25 V·cm. Between the lines: - The result is less about a single diagnostic model and more about a hardware shift: moving AI inference from electronics to photonics. - That shift matters because the bottleneck in medical AI is increasingly not model accuracy alone, but speed, heat, and energy cost at scale. - The industry–academia structure also matters, because the project moves from material discovery toward system integration and possible commercialization. - The researchers’ focus on linearity and bandwidth suggests the field is still working through scaling problems before photonic AI can handle more complex clinical workloads. What’s next: - The team plans to scale the platform with wavelength-division multiplexing. - Using the device’s roughly 30 nm modulation bandwidth and dense resonance comb, a 40-channel WDM system could raise computational density by about 40 times without increasing clock speed. - That scaling approach could enable about 4,000 multiplications per layer. - The researchers also plan stronger industrial encapsulation, including atomic layer deposition of Al2O3, and large-area chemical vapor deposition growth to improve stability and manufacturing consistency. - Future work is expected to push the platform toward clinical use and broader photonic AI applications in medical imaging, drug discovery, and genomic analysis. The bottom line: - Shenzhen University’s photonic AI platform shows that light-based computing can deliver clinically relevant diagnostics with major gains in speed and energy efficiency, but scaling and manufacturing remain the next hurdles.