Idea 20/Quantum & Photonics/Instrument proposal/2 min

Put the Reconstruction Inside the Scanner

MRI reconstruction is a fixed linear inverse problem. That is exactly what an optical processor does cheaply.

Photonic inference can move magnetic resonance reconstruction from a GPU farm to the scanner itself, cutting latency enough to make real-time adaptive scanning routine rather than a research protocol.

01What changed

Photonic AI hardware moved from demo to capital: Neurophos raised a 110 million dollar Series A in 2026 to launch exaflop-scale photonic AI chips, with claims of an order-of-magnitude advantage over GPU systems.1 Duke Capital Partners, 2026
Integrated photonic processors also demonstrated reconfigurable quantum gates in a time-multiplexed architecture, showing that programmable linear optics is no longer a laboratory curiosity.2 Nature Communications, 2026

02The gap

Reconstruction latency is treated as a compute problem solved by buying more GPUs. But the dominant operation is a multidimensional Fourier transform followed by a regularised inverse, both of which are linear and both of which optics performs at the speed of light with almost no energy.

03The idea

Build a photonic co-processor that implements the coil-sensitivity-encoded forward operator and its regularised inverse as a fixed optical network, with the regularisation parameter set electronically. The scanner streams raw k-space into the optical stage and receives images back within a reconstruction cycle.

The enabling insight is that the operator is static per protocol. A photonic mesh can be calibrated once for a given coil geometry and sequence, then reused for every scan, which sidesteps the programming difficulty that plagues general photonic computing.

Latency is the product, not throughput. Sub-second reconstruction enables real-time adaptive scanning: the operator sees the image forming and changes the next acquisition, which currently requires shipping data to a cluster and waiting.

The measurable claim is image quality at parity with a GPU pipeline, measured by structural similarity and radiologist preference, at an order of magnitude lower energy per reconstruction and under one second end to end.

04Why now

Photonic matrix hardware reached commercial scale-up in 2026, and MRI manufacturers are under pressure on both energy and throughput. The operator is unusually well suited because it never changes.

05The decisive test

Do this

A bench prototype reconstructing retrospectively acquired data for two sequences, benchmarked against a production GPU pipeline on quality, latency and energy.

Success looks like

Radiologist non-inferiority on 200 cases, reconstruction under one second, and energy per reconstruction at least five-fold lower.

Kills it

If calibration drift forces recalibration per scan, the fixed-operator advantage evaporates and the economics collapse.

06What would break it

Analog precision may be insufficient for diagnostic contrast requirements.
Integration into a scanner requires medical device certification of a novel compute element.
The same inverse problem is increasingly solved well by highly optimised low-precision digital hardware.

07If it works

Real-time adaptive MRI becomes standard, which shortens scans, reduces motion artefact and lowers the energy cost of a diagnostic pathway.