A tunable metasurface is a flat panel of small resonant elements whose radio response can be electronically retuned. Panels like this can image a 3D scene without moving parts. Each frequency and each tuning state projects a different scrambled radiation pattern into the scene, and a computer then inverts the measurements into a volume. This approach already works at human scale for metallic targets: a 2.1 m aperture of frequency-diverse panels reaches the 6.1 mm diffraction limit using only 96 switching channels (Large Metasurface Aperture for Millimeter Wave Computational Imaging at the Human-Scale). It has not been made to work inside soft tissue. Three things go wrong there. Waves lose energy on the way in and again on the way out, the same multiplicative "double fading" penalty that hurts passive relay surfaces in wireless links. Neighboring tissues differ only slightly in permittivity, so boundaries are faint. And the per-slice learned decoders used in microwave stroke imaging blur exactly those boundaries.
The core idea is to tune the panel and reconstruct the image as one system, using a physically correct model of how each element's response changes with frequency. Deeper tissue absorbs higher frequencies more strongly, so the useful bandwidth shrinks as depth grows. A physics-correct aperture model lets us choose tuning states that keep the surviving low-frequency measurements informative about deep, faint boundaries. A 3D decoder that knows depth and boundary uncertainty then turns those measurements into a volume. The bet is that this joint design, not a bigger network, is what recovers low-contrast structures at depth.
Core Claim
Each metamaterial element behaves like a driven oscillator with a Lorentzian response. In the form used for dynamic metasurface antennas (Dynamic Metasurface Antennas for 6G Extreme Massive MIMO Communications; Wideband Multi-User MIMO Communications with Frequency Selective RISs) it is
$$\alpha(\omega; \omega_r) = \frac{F\,\omega^2}{\omega_r^2 - \omega^2 + j\omega\Gamma},$$
where $\omega$ is the frequency, $\omega_r$ is the tunable resonance frequency, $F$ is the oscillator strength and $\Gamma$ is the damping. Under a first-order (Born) scattering approximation, the measurement at frequency $\omega$ and tuning state $c$ is
$$y(\omega, c) = \sum_v E_{t}(v,\omega,c)\,E_{r}(v,\omega,c)\,e^{-2\kappa(\omega) d_v}\,\Delta\varepsilon(v) + n,$$
where $E_t$ and $E_r$ are the transmit and receive fields that the Lorentzian elements produce at voxel $v$. The factor $e^{-2\kappa(\omega) d_v}$ is the two-way loss over tissue depth $d_v$ with frequency-dependent absorption $\kappa(\omega)$. The quantity $\Delta\varepsilon(v)$ is the permittivity contrast we want, and $n$ is noise.
The product $E_t E_r e^{-2\kappa d}$ is the imaging counterpart of double fading. Deep voxels sit at the bottom of the same multiplicative curve that mid-link relay surfaces suffer from (Mitigating Double Fading in Reconfigurable Intelligent Surface Networks). In imaging, though, the aperture cannot be moved to the endpoint, because the endpoint is inside the body.
Our claim has three parts. First, with the correct Lorentzian forward model, the tuning schedule $\{\omega_r\}$ can be optimized to maximize information about deep boundary voxels within the bandwidth that survives attenuation. Second, a 3D decoder that receives an explicit per-voxel depth and attenuation map, and that predicts boundary uncertainty, recovers faint inclusions better than per-slice decoders. Third, the frequency-correct model matters beyond what a flexible network can learn on its own. This third part is the one most likely to fail, and the experiments are built to test it directly.
Research Program
We first build a differentiable forward simulator. It contains a planar aperture of Lorentzian elements with tunable resonance, operating in a low-GHz band where tissue penetration is plausible. The scene is a lossy tissue-mimicking medium.
Ground-truth geometry comes from real anatomy. We take organ and lesion label maps from NIH Pancreas-CT and the Medical Segmentation Decathlon. We rescale them and embed them as dielectric inclusions at controlled depths and contrast levels. The point of this is not to claim microwave pancreas imaging. These labels supply exactly the irregular, low-contrast 3D shapes that defeat current segmenters, and they come with voxel-level truth.
The forward model starts in the Born approximation. A full-wave solver checks a subset of scenes so we know where the linear model breaks down.
Second, we optimize the tuning schedule. The tuning states are treated as learnable parameters. They are trained end to end with the decoder against a boundary-weighted reconstruction loss. This borrows the joint active/passive precoding idea from wideband reflecting-surface work (Efficient Joint Precoding Design for Wideband Intelligent Reflecting Surface-Assisted Cell-Free Network), but the objective is image information rather than sum-rate. The physical lesson that schedules optimized for one frequency compromise the band edges (Frequency Reflection Modulation for Reconfigurable Intelligent Surface Aided OFDM Systems; Beamforming Design for Wideband Near-Field Communications With Reconfigurable Refractive Surfaces) becomes, here, a trade-off. Energy can go to the high frequencies, which give fine resolution near the surface, or to the low frequencies, which survive to depth.
Third, we build the decoder. Its input is a physics-based back-projection volume. Its output is permittivity contrast plus a boundary-uncertainty map. It is a native 3D network in the spirit of UNETR and the volumetric state-space models. For the low-contrast margins it borrows the multi-scale deformable boundary path of DB-MSMUNet. That paper reported a 2D design and explicitly left through-plane consistency untested. We extend it to 3D.
Following DUCore, boundary supervision is weighted by predicted uncertainty. This guards against the known failure of edge branches that fit noise and speckle. We add one new ingredient: a per-voxel attenuation map $e^{-2\kappa d_v}$ fed to the decoder as a conditioning channel, so the decoder knows how much to trust the evidence at each depth.
Baselines
The naive reference is linear back-projection, or the direct sampling method, on the same measurements, with the same aperture and a uniform random tuning schedule. Any learned method must beat it. Its threshold-based segmentation gives the floor for Dice and for the 95th-percentile Hausdorff distance.
The strongest existing family for learned microwave tissue imaging is direct sampling followed by a 2D U-Net (Microwave Imaging of Conductors by Direct Sampling Method and U-Net), along with the field-enhancement U-Net for stroke (Deep Learning-Based Electric Field Enhancement Imaging Method for Brain Stroke). We found no reported result on Lacuna for either method on low-contrast dielectric 3D inclusions. We will reimplement both on our simulated data with conventional antenna arrays of matched element count and bandwidth.
The strongest generic segmenter is a 3D nnU-Net trained on back-projected volumes. On CT it reaches 0.94 DSC on liver but only 0.21 DSC on pancreas tumors (The Medical Segmentation Decathlon). The strongest low-contrast boundary model is DB-MSMUNet, which reports 89.47% DSC on NIH Pancreas from CT. We use that architecture, trained slice-wise on back-projected volumes, as the 2D boundary-specialist competitor. We also train it on the original CT slices of the same geometries. That CT run is an anatomical ceiling showing how far microwave measurements fall short of CT; it is not a target.
Fairness rules apply to every learned method. All see identical simulated measurement sets and train/test splits. All have the same GPU-hour budget and the same hyperparameter search budget, set at nnU-Net's automatic configuration plus an equal number of manual trials for the others.
The ablations isolate the proposal's ingredients one at a time:
- replacing the Lorentzian model with a frequency-flat phase model in both the simulator used for training and the optimization of the schedule, while keeping the Lorentzian model as ground truth for the measurements;
- replacing the optimized schedule with a random one;
- removing the attenuation-conditioning channel;
- running the decoder slice-wise.
Beating the strongest baseline means a statistically significant improvement, over five seeds, in Dice and in 95th-percentile Hausdorff distance for inclusions in the deepest third of the depth range and the lowest contrast bin. Gains on shallow, high-contrast inclusions do not count.
Why not just…?
Why not just use a conventional antenna array with a U-Net, as the stroke-imaging work does? This tests whether the metasurface earns its place. The stroke U-Net papers report blurry margins, but their arrays have many more RF chains. Our matched comparison fixes element count and bandwidth and reports performance against channel count. That sweep is where the frequency-diverse aperture's channel savings, shown for metallic targets in the human-scale aperture paper, should matter. If the conventional array wins at equal channel count, the hardware claim fails and we will say so.
Why not just model the panel as frequency-flat and let the network absorb the error? This is the most serious threat. A decoder trained on measured data can learn a fixed calibration offset. Our counterargument is narrow. The Lorentzian dispersion depends on the tuning state, so a network trained on one set of tuning schedules cannot learn the correction for schedules it has never seen. The schedule optimization also needs gradients through a correct model. The decisive test trains both the flat and the Lorentzian versions, then evaluates on held-out tuning schedules and on schedules optimized under each model. If the flat model plus a large network matches ours on held-out schedules, we will drop the physics claim and keep only the depth-aware decoder.
Why not just lower the frequency to avoid attenuation, or segment back-projections with nnU-Net? Lowering the frequency trades depth for resolution, which is the very trade-off the optimized schedule navigates. A fixed low band is therefore included in our schedule comparison as a special case. The two-stage nnU-Net pipeline is a baseline above. If it matches us, the result means the measurement design, not the decoder, is the bottleneck. That is still informative.
Proposed Experiments
The main experiment fixes a depth-by-contrast grid of inclusions derived from the MSD and NIH Pancreas geometries. At each cell it reports Dice, 95th-percentile Hausdorff distance, and a through-plane continuity score: the fraction of slice pairs whose connected inclusion components overlap. This last metric directly measures the slice-to-slice fragmentation documented in Deep convolutional networks for pancreas segmentation in CT imaging.
Three checks decide the claims. The first is a held-out-schedule test separating the Lorentzian and flat forward models. The second is a depth-resolved curve, which should show the optimized schedule's advantage growing with depth if the attenuation story is right. The third is a full-wave validation on a subset of scenes, to confirm that gains measured under the Born model survive multiple scattering. As a final sanity check on the conductive-target regime, the simulator should approach the reported 6.1 mm cross-range resolution when configured like the human-scale aperture. Without that agreement, the soft-tissue results cannot be trusted.
References
- Large Metasurface Aperture for Millimeter Wave Computational Imaging at the Human-Scale
- Dynamic Metasurface Antennas for 6G Extreme Massive MIMO Communications
- Wideband Multi-User MIMO Communications with Frequency Selective RISs: Element Response Modeling and Sum-Rate Maximization
- Pricing for Reconfigurable Intelligent Surface Aided Wireless Networks: Models and Principles
- The Medical Segmentation Decathlon
- Efficient Joint Precoding Design for Wideband Intelligent Reflecting Surface-Assisted Cell-Free Network
- Frequency Reflection Modulation for Reconfigurable Intelligent Surface Aided OFDM Systems
- Beamforming Design for Wideband Near-Field Communications With Reconfigurable Refractive Surfaces
- UNETR: Transformers for 3D Medical Image Segmentation
- DB-MSMUNet: Dual Branch Multi-Scale Mamba UNet for Pancreatic CT Scans Segmentation
- DUCore: Dual Uncertainty-Guided Consistency and Regional Contrastive Learning for Semi-supervised Medical Image Segmentation
- Microwave Imaging of Conductors by Direct Sampling Method and U-Net
- Deep Learning-Based Electric Field Enhancement Imaging Method for Brain Stroke
- Deep convolutional networks for pancreas segmentation in CT imaging