In remote and mobile consultations, a clinician often photographs a radiograph on a monitor or film with a phone and sends the photo onward, as mobile eHealth pipelines such as the cerebrovascular detection platform assume. We propose putting a thin patterned mask in front of the camera sensor. The sensor then records a small, scrambled set of measurements instead of a full image. Only a decoder tied to the exact position and rotation of the mask can turn those measurements into a correct diagnosis. So the device sends fewer bits, and the image never exists as plaintext on the phone. The difficulty is the one that holographic storage has fought for decades: optical decoding is extremely sensitive to position. A shift of a fraction of a pixel scrambles the signal (Menetrier and Burr). That sensitivity is what makes the mask work as a key. It is also what makes ordinary mechanical drift break authorized decoding.
Our central claim is that one designable property of the mask, the width of its autocorrelation, controls three things at once: how much drift the authorized user tolerates, how many distinct keys one physical mask provides, and how much high-frequency diagnostic detail survives compression. Autocorrelation width is the distance over which a shifted copy of the pattern still resembles the original. We propose to train the mask and a diagnostic decoder jointly so that diagnostic accuracy stays flat inside a chosen drift window and collapses outside it. We would then measure, on chest radiographs and mammograms, the trade-off between key count, drift tolerance and diagnostic accuracy at a fixed number of transmitted bits. As far as the file shows, no prior work treats optical misalignment as a quantity to shape in this way rather than to maximize or to eliminate.
Core Claim
Existing optical keying work pushes sensitivity as high as possible. Xu et al.'s dual-path encryption system uses a partially coated beam splitter to route an important region through double random phase encoding. A ±6-pixel shift of the splitter makes the correlation coefficient collapse. Minamata et al.'s coded aperture key reaches 82.1% character recognition with the correct aperture, against 7.7% with a circular one. OpEnCam builds a deliberately shift-variant two-mask lensless camera and decodes it with a keyed transformer. None of these systems reduces the transmitted data rate. None was tested on diagnostic tasks. None asks what happens when the authorized device itself drifts, which the research file lists as an open deployment failure.
Holographic storage has the opposite goal: it tries to remove sensitivity, and it pays for doing so. Menetrier and Burr show that digital shift compensation recovers pages shifted by nearly half a pixel, but only by amplifying noise and giving up density.
We treat sensitivity as a function to be designed. Let $y_d = S\,(P_\theta \ast x)_d + n$ be the $m$ sensor readings of image $x$ through mask $P_\theta$ displaced by $d$. Here $S$ is sensor sampling and $n$ is noise, the same forward model OpEnCam uses. We train mask parameters $\theta$ and decoder $f_\phi$ with
$$\min_{\theta,\phi}\; \mathbb{E}_{d\sim W}\,\ell\big(f_\phi(y_d),c\big)\;-\;\lambda\,\mathbb{E}_{d\notin W_\gamma}\,\ell\big(f_\phi(y_d),c\big),$$
where $c$ is the diagnostic label, $\ell$ is cross-entropy, and $W$ is the drift window the device must tolerate (for example ±1 sensor pixel and ±0.5°). $W_\gamma$ is $W$ widened by a guard band $\gamma$, and $\lambda$ weights the penalty for succeeding off-key. The second term is capped so it cannot drive the decoder toward pathological outputs.
The key is the offset $d$ that the device applies on purpose each session, with the mask mounted on a small piezo or rotation stage, or displayed on a transmissive liquid-crystal panel. We use rotation because Fan et al. showed that a twist angle can index hundreds of independent diffractive states. The number of usable keys is roughly
$$K \approx \Big(\frac{A}{w+\gamma}\Big)^2 \cdot R,$$
where $A$ is the translation range, $w$ is the autocorrelation width of the mask, and $R$ is the number of separable rotations. The form follows from counting non-overlapping decorrelation cells. Drift tolerance requires $w$ to be larger than the drift. Preserving fine detail such as microcalcifications pushes toward fine mask structure, and so toward small $w$. This tension is the one the file's Gap 3 identifies: Zhang et al. found that optical encoders favor low frequencies, while mammography diagnosis depends on high-frequency detail (Kar et al.). The research question is where this three-way frontier lies for real diagnostic tasks.
Research Program
We would first build a differentiable simulator of the mask-plus-sensor forward model. It would include sub-pixel translation, rotation, defocus, and Poisson-plus-Gaussian noise. The simulator has two roles: training, and sweeping the misregistration–accuracy curve that the file's Gap 2 says no one has measured.
The mask is parameterized as a smooth random field whose correlation length $w$ is an explicit, sweepable hyperparameter. The remaining pattern is learned.
The decoder does diagnosis directly, not image reconstruction, because the file's baseline methods argue that diagnostic information is sparse. We also train a reconstruction head so that radiologist-style image quality can be reported.
Following the physics-in-the-loop idea of Minamata et al., we would then build a bench prototype. It consists of a phone-class sensor, a printed or liquid-crystal mask, and a translation/rotation stage, pointed at a calibrated medical display. The bench is used to fine-tune the simulator's aberration model. This is an emulation of the phone-photographs-monitor workflow. It is not a claim about X-ray detectors.
Baselines
The naive reference is the plain phone photograph compressed losslessly with High-Throughput JPEG 2000 (HTJ2K) in the manner of MIST, and then encrypted with AES. MIST reports a 56.22% average transmission reduction losslessly and 96.10% at its lowest sub-resolution, with mean SSIM 0.83. We run the same codec at matched bit budgets, from lossless down to our measurement budget.
The strongest diagnostic-rate competitor is software pruning in the style of NEURAL. It reports a 97.7% data reduction on MIMIC-CXR with pneumonia detection AUC of 0.88–0.95, using report-guided attention. We run it with the same backbone family and the same tuning budget as our decoder. We report it both with reports, which is its intended setting, and with reports withheld, which is the realistic triage setting.
For mammography, the learned codec of Kar et al. reports compression beyond 300× (below 0.04 bits per pixel) on CBIS-DDSM. We retrain it and attach a diagnostic head of the same capacity as ours.
The optical baseline is spatial binning and the orthogonal-point-spread-function encoder of Zhang et al.. The file reports no numbers for these ("no reported result found on Lacuna" beyond comparative curves), so we reimplement them at equal measurement count.
Two ablations isolate our contribution. The first is the same learned mask trained without the drift-window term, which amounts to an OpEnCam-style maximally sensitive key. The second is that same mask paired with Menetrier–Burr-style digital shift compensation.
Beating the baselines means the following. At equal transmitted bits and with drift inside $W$, our AUC must be within 0.02 of the best digital baseline. Off-key AUC must be near chance. Neither ablation may achieve both properties at the same time.
Why not just…?
Why not just use HTJ2K or HEVC plus AES on the phone? This is the real threat. On diagnostic accuracy per bit, a good digital codec will probably match or beat us. Our case rests only on two things: the plaintext never reaches phone memory, which is OpEnCam's argument, and the sensor needs fewer readout operations. If the experiments show no measurement-count or energy advantage at equal AUC, the optical approach is not justified for this workflow, and we would report that.
Why not just use NEURAL-style pruning? It needs a paired report to decide what to keep, which does not exist at the moment of triage. Its output is also unprotected. The report-withheld run of NEURAL tests the first point directly.
Why not just maximize sensitivity and correct drift in software? Menetrier and Burr show that compensation amplifies noise and gives up density. The second ablation measures whether this costs diagnostic accuracy at our bit budgets. If it matches our jointly trained mask, the drift-window term is unnecessary, and we would drop it.
Decisive Checks
We run everything on MIMIC-CXR (pneumonia and effusion labels) and on CBIS-DDSM calcification cases. We sweep $w$ and measure diagnostic AUC as a function of displacement, both in simulation and on the bench.
The first decisive check is whether a single mask can give both a flat AUC plateau inside ±1 pixel and near-chance AUC beyond the guard band, while keeping mammographic calcification sensitivity. Calcification sensitivity is where Gap 3 predicts failure.
The second decisive check is a known-plaintext attack. The attacker receives $n$ paired images and measurements at one session key, then trains a fresh decoder. We report the $n$ at which the attacker's AUC recovers. We compare that $n$ with how often keys are rotated, because it determines whether the keyspace $K$ gives any real protection.
References
- The Architecture of an Automatic eHealth Platform With Mobile Client for Cerebrovascular Disease Detection
- Density implications of shift compensation postprocessing in holographic storage systems
- Local hybrid optical encryption system based on double random phase encoding
- A Coded Aperture as a Key for Information Hiding Designed by Physics-in-the-Loop Optimization
- OpEnCam: Lensless Optical Encryption Camera
- Holographic multiplexing metasurface with twisted diffractive neural network
- Convolutional Optical Encoders for Generalizable Image Compression
- Fully Convolutional Model for Variable Bit Length and Lossy High Density Compression of Mammograms
- Towards Resource-Efficient Streaming of Large-Scale Medical Image Datasets for Deep Learning
- NEURAL: Attention-Guided Pruning for Unified Multimodal Resource-Constrained Clinical Evaluation