How Many Bits Does a Virtual Biopsy Need?

Models that predict a tumor's genetics from MRI often look excellent at the hospital where they were built and collapse at the next one. The clearest case is the BraTS 2021 radiogenomics challenge. There, the winning model for predicting MGMT promoter methylation (a chemical tag on a DNA-repair gene that changes how glioblastoma responds to chemotherapy) fell to an AUROC of 56.2% on an independent hospital cohort, and 80.2% of 420 models were no better than chance (Validation of MRI-Based Models to Predict MGMT Promoter Methylation in Gliomas: BraTS 2021 Radiogenomics Challenge). A leading explanation is that fine-grained voxel intensities carry two kinds of signal. One is coarse biological structure, such as necrosis, enhancement patterns, and edema. The other is a fine-grained fingerprint of the scanner, coil, and reconstruction software. Models trained at one site learn both, and only the first kind travels.

We propose to measure this directly by borrowing a result from wireless engineering. Reflecting surfaces whose elements can take only 1 or 2 bits of phase lose almost nothing compared with continuous control, because quantization errors spread over many elements average out. Past 2–4 bits, extra precision buys little and costs power. We hypothesize the same saturation in radiogenomics. Information about the tumor's genotype should saturate at a few bits of intensity resolution per modality, while information identifying the acquisition site should keep growing with resolution. If that holds, a learned, coarse, scanner-invariant quantization of the images, chosen to maximize genotype information per bit rather than raw internal accuracy, should narrow the gap between internal and external performance. We would lock it at derivation and test it on held-out cohorts.

Core Claim

The claim is empirical and falsifiable. Take a fixed radiogenomic pipeline and sweep the intensity resolution $b$ (bits per modality) from 1 to full precision. Two curves come out of this sweep. The first is externally validated genotype AUC, which should rise and then plateau at small $b$. The second is how well a simple probe can identify the acquisition site from the same features, which should keep rising with $b$. The gap between them is the region where extra resolution mostly buys scanner fingerprint. A second, more constructive claim follows from the first. A model that quantizes intensities against per-scan quantile thresholds, learned jointly with the classifier, and that sparsely activates only a few modality-by-region "modules," should beat continuous-intensity models with standard harmonization on external cohorts. It should do so for a target with real imaging signal (IDH). For MGMT, it should at least reveal whether any transportable signal exists at all.

Prior work contributes three different things to this claim. The radiogenomics and validation literature motivates it. The BraTS/SNUH failure (art_ab403e) and the prostate radiomics review finding that radiomic features are highly sensitive to vendor, field strength, and reconstruction (Radiomics in prostate cancer imaging for a personalized treatment approach) show the gap but never relate it to representation resolution. The discrete-phase literature supplies the mechanism we test. On the Optimality of Reconfigurable Intelligent Surfaces (RISs) proves that 1-bit control preserves the coherent $O(N^2)$ gain. Energy Efficient Multi-User MISO Communication using Low Resolution Large Intelligent Surfaces finds that 2 bits match continuous performance. Noise-Robust One-Bit Diffraction Tomography and Optimal Dose Fractionation shows that 2- and 4-phase masks match continuous masks and that 1-bit thresholding stays stable up to a noise-to-signal ratio of 0.5. That last result comes only from a 36³ synthetic phantom, so it suggests rather than establishes a denoising effect on real tissue. The energy-efficiency literature supplies the evaluation lens. Rate over power, $R/(P_{\text{static}} + N P_{\text{element}})$, peaks at an optimal surface size (Reconfigurable Intelligent Surfaces for Energy Efficiency in Wireless Communication). Activating a sparse subset of surface modules beats switching all of them on (Reflection Resource Management for Intelligent Reflecting Surface Aided Wireless Networks). We reuse that module-sparsity idea for modality and region selection.

Research Program

The method has three parts. First, a quantizer. Each modality $m$ of each scan is mapped to $2^{b_m}$ levels using thresholds set at learned quantiles of that scan's own intensity distribution inside a brain mask. Because the thresholds sit on within-scan ranks rather than absolute intensities, the quantizer is invariant to any monotone intensity remapping, which is the dominant form of between-scanner shift. The thresholds are trained with a straight-through gradient. The bit depth $b_m \in \{1,2,3,4,\text{full}\}$ is chosen per modality by greedy coordinate descent, in the spirit of the $O(N)$ discrete selection algorithm in art_094f4e. Second, a modular feature extractor. Features (texture statistics plus a small 3D CNN) are computed per modality-by-region module: enhancing tumor, necrosis, and edema from an nnU-Net segmentation, as in Deep-learning and conventional radiomics to predict IDH genotyping status. A mixed $\ell_{1,F}$ group penalty, taken from reflection resource management, switches whole modules off. Third, a saturation-penalized objective chosen on derivation data only:

$$ \max_{\theta,\,b}\;\; \frac{\text{AUC}_{\text{LOSO}}(\theta,b) - 0.5}{c_0 + \sum_{m \in \mathcal{A}} b_m} $$

In this objective, $\theta$ is the set of model weights and $b$ is the vector of per-modality bit depths. $\text{AUC}_{\text{LOSO}}$ is the AUC under leave-one-site-out cross-validation within the derivation pool. $\mathcal{A}$ is the set of active modules, and $c_0$ is a fixed cost per active module, playing the role of static circuit power. The form copies the bits-per-Joule ratio, so it penalizes resolution that buys only within-site gains. After selection, we lock all parameters, thresholds, bit depths, and the decision threshold before any external cohort is touched, following the derivation-and-validation discipline.

The analysis component is the site-information probe. At each $b$, a logistic regression is trained to predict acquisition site from the frozen features. We plot site-identification AUC against genotype AUC across the sweep. This plot is the central scientific output, whether or not the constructive method wins.

Baselines

The naive reference is chance, meaning external AUROC 0.5, along with the full-precision version of our own pipeline with no harmonization. The strongest existing external reference for MGMT is the BraTS 2021 first-place model, which reached AUROC 56.2% and accuracy 54.8% on SNUH (art_ab403e). We will rerun it from its released form on exactly our external splits rather than quote it. For IDH, the strongest pipeline in the file is the nnU-Net plus eca_nfnet_l0/Bi-LSTM plus logistic regression model with AUC 0.920 under nested 10-fold cross-validation (art_70272a). That number is internal, and no locked external figure is reported in the file. We will reimplement the pipeline, train it on the same derivation pool, and report its external AUC as the bar.

The harmonization baseline is the same pipeline at full precision with z-score normalization and ComBat feature harmonization. It gets an identical feature budget and hyperparameter search. The ablation that isolates our key ingredient is fixed, uniform gray-level binning at each $b$, which is the standard radiomics discretization, compared against our learned quantile quantizer at the same $b$. A second ablation removes the group penalty. All models share the segmentation, the derivation data, and an equal tuning budget of 200 configurations under leave-one-site-out cross-validation. Beating the field means two things on locked external cohorts. For IDH, the external AUC must exceed both the reimplemented 0.920-internal pipeline and ComBat, with a DeLong test at p<0.05. For MGMT, the external AUROC must exceed 0.5 with a confidence interval that excludes chance. Merely beating 56.2% would not count.

Why not just…?

Why not just tune the gray-level bin count, which radiomics already does? This is the most serious threat, and it is exactly the fixed-binning ablation. If uniform binning at the best $b$ matches the learned quantizer externally, then the constructive method reduces to a practical recommendation: use few bins and choose them by cross-site validation. The saturation analysis would still stand, because nobody in the file has measured site information against biological information as a function of resolution. We expect quantile thresholds to help, because uniform bins over absolute intensity are not invariant to scanner-specific intensity remapping. But this is an expectation, and the experiment will decide it.

Why not just harmonize with ComBat and keep full precision? ComBat corrects feature means and variances per site, but it needs site labels at test time and cannot remove nonlinear texture fingerprints. The BraTS evidence shows that standard preprocessing combined with strong CNNs did not transport for MGMT (art_ab403e). ComBat runs as a full baseline. If it closes the gap as well as quantization does, we report that.

Why not just recalibrate locally at each site, as All models are local argues? Local recalibration fixes miscalibration but cannot create discrimination that the features lack. It also requires labeled local data, which for MGMT means tissue. We include a recalibrated variant of each model at each external site. If recalibration recovers AUC, which it cannot do through a monotone transform of a single score, our premise is wrong.

Why not change the acquisition physically, using programmable metasurfaces or phase-coded apertures in the scanner? That is the literal bridge between the four directions, but no raw k-space data or hardware exists in any cohort in the file. The one quantized-mask tomography result is purely synthetic (art_8a81c5). We test the cheaper and prior question of whether coarse discrete representations carry the transportable signal. A positive answer would justify the hardware work, and a negative one would save it.

Proposed Experiments

The derivation pool is BraTS 2021 ($n=585$), split by contributing institution where institutional metadata can be obtained. External test cohorts are SNUH ($n=400$, MGMT, requiring a data agreement with its authors) and, for IDH, TCGA-GBM/LGG ($n=218$) used as the locked external cohort, with Nanjing ($n=268$) if access is granted. The IDH experiment is the positive control, because its perfusion and structural signal is biologically grounded. If quantization hurts IDH externally, the hypothesis fails. A secondary replication uses the dual-center rectal Ki-67 setting, where a combined radiomics model reached external AUC 0.86 at Center 2 (A novel radiomics based on multi-parametric MRI for predicting Ki-67 expression in rectal cancer). That model is the bar there if the data can be obtained.

Three checks decide the outcome. First, in the saturation plot, external IDH AUC should plateau by $b \le 3$ while site-probe AUC keeps rising to full precision. If both curves saturate together, coarse quantization trades away biology and scanner signal equally, and the claim is refuted. Second, the gap between internal leave-one-site-out AUC and external AUC should shrink monotonically as $b$ decreases toward the selected optimum. Third, the locked learned quantizer should beat fixed binning and ComBat externally on IDH. For MGMT, we report honestly whether any configuration in the sweep exceeds chance externally. If none does, that would be evidence against a transportable conventional-MRI MGMT phenotype, which is useful in its own right. We also report calibration, decision-curve net benefit, and bootstrap confidence intervals for every external number.