Does Synthetic ADC Make Radiologists Agree for the Wrong Reason?

Deep networks can now turn one kind of MRI scan into another. A prominent example turns the routine anatomical prostate scan (T2-weighted MRI) into a synthetic version of the apparent diffusion coefficient (ADC) map. The ADC map shows how freely water moves through tissue, and dense tumors restrict that movement. Radiologists prefer these synthetic maps because they lack the gas and motion artifacts that spoil real diffusion scans. But the synthetic map is computed entirely from the T2 image the radiologist is already reading. As a matter of basic information theory, it cannot contain any evidence about the tumor that T2 lacks. It can only re-display T2 evidence, or add plausible-looking content that is not really there.

This note proposes a reader study and a machine-reader study built around that constraint. We ask what happens to disagreement between radiologists when a synthetic map replaces the acquired one. If readers agree more but are no more accurate against biopsy, the synthetic map is manufacturing consensus rather than revealing disease. We test this across hospitals and in a silent prospective stream. We concentrate the test on the lesions where the two physical scans disagree, because that is where a T2-derived map must fail if it fails anywhere.

Core Claim

Let $X$ be the T2 image, $A$ the acquired ADC map, $\hat A = g(X)$ the synthetic map from generator $g$, and $Y$ the biopsy-confirmed presence of clinically significant cancer. Because $\hat A$ is a function of $X$ alone, the data processing inequality gives

$$I(Y; X, \hat A) = I(Y; X) \le I(Y; X, A),$$

where $I$ is mutual information. A reader with unlimited computation gains nothing from $\hat A$. A human reader can still gain, because humans cannot compute $g$ in their heads. Any benefit must therefore come from making existing T2 evidence easier to see.

Two consequences can be tested. First, the synthetic map should help least, and possibly hurt, on lesions whose diffusion restriction is visible on $A$ but not reflected in $T2$ texture. Second, a rise in agreement between readers is not evidence of validity. Suppose every reader's decision threshold shifts in the same direction because they all see the same smooth synthetic image. The Fleiss' kappa statistic, which measures agreement across many readers beyond chance, will then rise while accuracy stays flat. We call this "manufactured consensus." Our main contribution is to measure it directly, by separating each reader's systematic bias from their random noise.

The motivating evidence comes from the AI-ADC multi-reader study. It trained a contrastive, attention-based T2-to-ADC generator. Four readers preferred its output over acquired ADC in 34.80% versus 5.07% of cases and reported a sharper prostate edge ($\beta = 1.38$). But the study covered 74 single-center cases and explicitly did not measure cancer detection. We take its generator design as the object under test and do not propose a new generator.

A second study shows why rising agreement is ambiguous. Kottlors et al. report AI assistance raising Fleiss' kappa from 0.34 to 0.58 alongside a modest accuracy gain, from 79.1% to 81.5%. That is a case where agreement and accuracy moved together. Our question is whether synthetic imaging decouples them.

Research Program

Generators. We train two T2-to-ADC generators on bi-parametric prostate MRI from PI-CAI, which has both T2 and acquired ADC. The first reimplements the contrastive attention design of AI-ADC. The second is the paired Pix2Pix model with a ResNet backbone and 2.5D input that performed best in Lapaeva et al., where it reached 63.21 HU mean absolute error for MRI-to-CT synthesis. We then apply both, frozen, to PROSTATE158 as an external center. The external test matters for two reasons. Bejar et al. showed that features sensitive to slice thickness collapse under leave-center-out testing. And Lapaeva et al. showed that voxel similarity metrics fail to predict small, clinically relevant errors. For both reasons we report voxel metrics but do not treat them as evidence of safety.

Discordance stratification. Before any reader sees a case, a radiologist not in the reader pool labels each index lesion from the acquired images alone. The label records whether the lesion is conspicuous on T2, conspicuous on acquired ADC, both, or neither. The key stratum is "ADC-conspicuous, T2-subtle." The data processing argument says a T2-derived map carries no information about these lesions beyond T2 itself. We enrich this stratum in the reader set so the study has power where the risk sits.

Human reader study. Six radiologists read each case in three arms, separated by washout periods and in randomized order, following the crossover design of Kottlors et al.. The arms are T2 alone, T2 plus acquired ADC, and T2 plus synthetic ADC. Readers give a lesion-level PI-RADS-style score and a biopsy recommendation.

Separating bias from noise. We fit a rater-conditioned latent model in the style of the Latent Doctor Model and Pionono. These methods place each annotator in a latent space whose mean captures systematic bias and whose variance captures inconsistency. We fit one such model per arm. Two outcomes are possible when the synthetic arm raises kappa. If the gain comes from shrinking reader variance, readers have become more consistent. If it comes from all reader means shifting together toward "benign" or "malignant," that is manufactured consensus. Accuracy against biopsy then tells us whether the shared shift was toward the truth.

Machine reader. The same three arms are given to a classifier. We use a small-data geometric classifier in the style of MFM-Geom and a plain convolutional network. This isolates the part of any human gain that is purely perceptual. The inequality predicts no gain for a machine from adding $\hat A$, except through regularization in small samples, which we measure by varying training size.

Silent prospective stream. At one partner site, the generator runs in shadow mode on consecutive prostate MRI exams. Its output is not shown clinically, but it is logged. We then compare it with acquired ADC and with eventual pathology. This exposes the uncurated case mix that Wang et al. and Zhao et al. argue retrospective sets hide, including missing sequences and borderline scans. We deliberately stop short of a live arbitration design like Ng et al.. Such a trial would be premature until the retrospective checks are passed.

Baselines

The do-nothing reference is T2 alone. By the argument above, it has the same information content as the synthetic arm. The synthetic map therefore has to beat T2 alone in human accuracy, not just in preference ratings, to justify its use. Readers and the machine classifier see identical cases, resolution, and reading-time limits in every arm.

The clinical standard is T2 plus acquired ADC, which is ordinary bi-parametric reading. It is the strongest competitor because it is the only arm that carries genuinely new physical information. Over the same case set, we treat the synthetic arm as acceptable only if it meets both of two conditions. First, it must be non-inferior to acquired ADC in biopsy-referenced AUC on the whole set. Second, it must not be worse than T2 alone in the ADC-conspicuous, T2-subtle stratum. We will pre-register the non-inferiority margin.

The strongest reported machine result in the file is MFM-Geom. It reports AUC-PR of 90.67 on PI-CAI when trained on 10% of the data, and 90.61 on external PROSTATE158 for its unimodal bi-parametric variant. That figure uses acquired ADC. We rerun the same architecture with the same training budget and tuning search in all three input arms. The question is not whether we exceed 90.61. It is whether swapping acquired for synthetic ADC closes, or fails to close, the gap to the T2-only run.

For the synthesis method itself, AI-ADC reports only perceptual outcomes. There is no reported diagnostic-accuracy result on Lacuna for AI-ADC.

The key ablation is a non-generative display transform of T2. This is a fixed windowing-and-inversion rendering matched in visual style to ADC. It isolates whether any human benefit comes from the learned, T2-derived content of the synthetic map or simply from seeing T2 rendered differently.

Why not just…?

Why not just trust the reader preference data already published? Preference is not detection. The AI-ADC study itself declined to measure diagnostic accuracy. Lapaeva et al. showed that image quality measures miss clinically relevant errors within tight tolerances. A sharper edge could be a hallucinated edge, and only biopsy-referenced accuracy in the discordant stratum can tell the two apart.

Why not just report Fleiss' kappa across arms? A rise in kappa is exactly the ambiguous signal this proposal targets. Common bias shifts and variance reduction both raise agreement. The latent rater models of Pionono and the Latent Doctor Model separate bias from inconsistency, and we pair that decomposition with biopsy accuracy. Kappa alone would score manufactured consensus as success.

Why not just show readers a better-rendered T2 image instead of a generated map? This is a real threat. If the display-transform ablation matches the synthetic arm in accuracy and agreement, the generator adds nothing beyond visualization, and we will report that. The design settles the question directly, because both arms are read by the same readers on the same cases.

Why not just give readers a T2-only cancer classifier's probability score? A classifier score also re-presents T2 information. Yuan et al. and Kottlors et al. show such assistance can raise accuracy and speed. It is a plausible alternative, so we run it as a fourth arm on a subset of cases. If a scalar score equals the synthetic map, the imaging form is unnecessary.

Why not just repeat bad diffusion scans? Repeats cost scanner time. The AI-ADC readers reported a reduced need to re-scan ($\beta = 2.23$). Our prospective log measures how often acquired ADC is non-diagnostic, which bounds how much a synthetic substitute could possibly matter.

Decisive Checks

The central prediction concerns the ADC-conspicuous, T2-subtle stratum. There, human sensitivity with the synthetic map should be no better than T2 alone and clearly below acquired ADC. If instead the synthetic map matches acquired ADC in this stratum, then either the generator is extracting T2 cues that humans miss, or our discordance labels are unreliable. We would test the latter by having a second, independent radiologist label the stratum.

The second check is the decomposition of any kappa gain in the synthetic arm. If the gain comes mostly from a common shift in reader means, with no gain in biopsy-referenced AUC, that confirms manufactured consensus.

The third check is transfer. On PROSTATE158, we measure whether the correlation between synthetic and acquired ADC, and the synthetic arm's reader accuracy, degrade more than those of the acquired arm. A larger degradation would mean the generator is tied to the training site's scanners and protocols, in the way Bejar et al. found for slice-thickness-sensitive features.

The fourth check comes from the silent prospective log. In consecutive cases with biopsy, we look for how often synthetic ADC looks unremarkable where acquired ADC showed restriction. This is the rate at which a live deployment would have hidden a tumor.