The Measurement, Not the Model: Making Neural CT Compression Provable

We built a neural codec that compresses full-dose CT 33% to 46% smaller than JPEG-2000 at the same diagnostic quality, with radiomics agreement of ICC 0.992 against the lossless original. That is a strong result. But the finding that will outlast the codec is a discovery about measurement itself: the standard way to check whether a compressed scan still supports diagnosis is dominated by artifacts of the measuring tool, not by the codec under test.

We isolated three distinct artifacts, each able to flip a verdict on its own, and built a validation harness that neutralizes all three. Under that harness, the codec's advantage holds. The harness is the part worth carrying into every future codec and every future modality.

Key Takeaways - The result: a neural CT codec beats JPEG-2000 by 33% to 46% at diagnostic quality, with radiomics-ICC 0.992 versus 0.918 and segmentation Dice 0.962 versus 0.864. - The discovery: diagnostic-equivalence validation is dominated by measurement artifacts.

Three distinct ones can each manufacture a wrong verdict about a codec that is actually sound. - The rule that matters most: a diagnostic-equivalence delta is uninterpretable without the measuring tool's own noise floor.

A fast segmenter that disagrees with itself by 0.04 can invent a regression that is not there. - The payoff: a four-part harness that makes neural medical-codec claims trustworthy, and a thickness-gated routing map where sub-millimeter CT earns a 3D model at 12.5% fewer bits while everything else routes to 2D.

Why medical CT is a hard place to compress

Medical CT is a hostile target for lossy compression. A single study is large, it is retained by law for years or decades, and it cannot be visibly degraded without clinical risk.

The whole appeal of a better codec is storage saved at hospital scale, which means the savings only count if the diagnosis is provably preserved. So the question is never "does it look fine." The question is "does a downstream clinical task, run on the reconstruction, reach the same answer it reached on the lossless original."

That framing sets a high bar and a specific one. We are not chasing a pretty image. We are chasing agreement with lossless on the measurements that actually drive reads: radiomics feature stability, and organ or lesion segmentation.

The bar: beat lossless, then beat diagnostic-quality classical codecs

On real lossless CT (LIDC-IDRI), the strong classical codecs already compress by roughly a factor of three: JPEG-XL at 2.985x, JPEG-2000 at 2.866x, JPEG-LS at 2.849x. That is the floor any learned model has to clear before it earns a look.

Below lossless, the interesting question is the near-lossless diagnostic band: how few bits can you spend before a clinical task starts to disagree with the original.

We mapped that frontier on classical codecs first, and it has a hard edge. JPEG-LS near-lossless holds its worst radiomics feature above ICC 0.995 down to about 2.1 bpp. Push to 0.78 bpp and JPEG-2000 collapses that worst feature to ICC 0.49, and a thin structure like a rib drops out of the segmentation entirely (Dice 0.0). That edge, not the average, is the thing to protect.

So the target was a dual gate: beat the classical codec on bits, and match the lossless original on both task gates, at a bit-rate below where classical starts to break (under about 2.1 to 2.3 bpp).

The core discovery: BD-rate is not diagnostic equivalence

Here is the insight the whole program turns on. BD-rate, the standard way to say "fewer bits at equal quality," measures bits against a fidelity proxy like PSNR. It says nothing, on its own, about whether a radiologist-facing task survives.

A model can post an excellent BD-rate and still shift a radiomics feature or shave a segment. BD-rate is necessary and never sufficient, and every routing decision has to clear the clinical task gates in addition.

The moment you take that seriously and start measuring diagnosis directly, you inherit a new problem: the noise of whatever tool measures diagnosis. That is where the three artifacts live.

Three measurement artifacts that flip neural-codec verdicts

This is the part that generalizes beyond our codec. Validating diagnostic equivalence, we isolated three distinct measurement artifacts, each capable of producing a confident, wrong verdict about a codec that is actually sound. Any team validating a neural medical codec will meet at least one of them.

Artifact one: the segmenter's own noise floor

The most dangerous, because a false result here looks clean. A 3D inter-slice model appeared to regress segmentation agreement against 2D (Dice 0.856 versus 0.883), a decisive-looking win for 2D. It was an illusion of the instrument.

The measurement used a fast, non-deterministic segmentation mode on two volumes, and that mode disagrees with itself by about 0.04 on the identical input, larger than the difference under test.

Re-run with the accurate, deterministic mode (self-agreement 0.9999 to 1.0000), reproduced at n=8 (0.958 versus 0.953), and scaled to n=16 on full volumes, both models land near 0.955 and 3D holds.

The rule: a diagnostic-equivalence delta is uninterpretable until you have measured the segmenter against itself. If the tool cannot reproduce its own answer, it cannot resolve your codec.

Artifact two: volume truncation

Fixed-size 128-slice sub-volume crops truncate organs at the crop boundary, and a truncated organ segments worse for reasons that have nothing to do with the codec.

On crops, both models dipped below the 0.95 floor (0.941 and 0.935). On full volumes both recovered to about 0.955. The depression was 0.01 to 0.02 of Dice, pure windowing. Validate on whole organs, not crops.

Artifact three: unbracketed BD-rate extrapolation

A BD-rate computed against a reference grid that does not bracket the codec's rate-distortion curve has to extrapolate beyond the data, and extrapolation can land on any sign.

The same low-dose volume reads plus 38% worse, minus 25%, or minus 28% depending only on which grid you pick. Only the bracketing grid is trustworthy, and it says the codec is about 28% cheaper. A single unbracketed comparison nearly sent us building a specialist model to close a gap that did not exist. Enforce a bracketing check before trusting any BD-rate.

Three artifacts, one lesson: in each case the codec was sound and the measurement was not. That is the discovery worth carrying into the next codec and the next modality.

The physics that holds up: 3D benefit is thickness-gated

Not everything is an artifact, and one real finding recurred independently in both the classical and the neural regimes, which is how you know it is physics. The benefit of a 3D, inter-slice model over a per-slice 2D model grows as slices thin.

Adjacent slices differ by about 41.7 HU mean-absolute at 2.5 mm spacing, correlated but far from redundant, so 3D barely helps or even hurts there (minus 4.6% for the neural model on thick slices). At mid spacing 3D saves 7.7%, still below our 10% routing bar. Only sub-millimeter slices, where neighbors are nearly redundant, clear the bar, at 12.5% fewer bits.

Neural 3D advantage over 2D by slice thickness. The benefit grows as slices thin; only sub-millimeter clears the 10% routing bar.

The advantage over the classical codec has its own honest shape. It is largest at low, preview-grade quality (83% fewer bits at 42.8 dB), where any efficient model shines, and it narrows as quality climbs: 46% at diagnostic quality (49.5 dB), 33% at deep near-lossless (54.7 dB). We certify at the hard end, near lossless, not at the flattering end.

What we ship: a routing map, not one codec

The output is a routing table, not a single codec. Sub-millimeter CT routes to the 3D inter-slice model, 12.5% cheaper in bits at genuinely equivalent diagnosis. Mid and thick CT route to the 2D model, because the 3D benefit there (7.7% and 4.6%) falls below the 10% bar and does not justify the added complexity. Low-dose and screening CT route to the same 2D model, about 28% cheaper than JPEG-2000 once measured correctly. Full-dose CT generally runs the 2D model at conservative settings (around 52 dB and above); the most aggressive setting we tested fails the segmentation floor (Dice 0.926 below 0.95), so we do not ship it. Each route is certified on the clinical criterion, not just on bitrate. This is the thesis in miniature: best model per instance, chosen by a criterion that survives scrutiny.

The harness: a trustworthy bar for diagnostic equivalence

The reason we stand behind the routing map is that every verdict cleared the same four-part harness, and we offer it as a standard for validating any neural medical codec:

1. Accurate, deterministic segmentation, never fast mode.
2. Full volumes, never truncated crops.
3. A segmenter self-agreement check on the lossless volume, so the metric's noise floor is known before any delta is read.
4. n of at least 16, at least two scanner vendors, leakage-guarded.

Skip any one and you risk certifying noise as signal. Run all four and a diagnostic-equivalence claim becomes something a reviewer, a regulator, or a hospital can trust. This is what makes future results, ours and the field's, worth believing, and it is why we are publishing the method alongside the number.

Honest limitations

These results are directional-to-cohort evidence, not population claims. The full-dose held-out set is n=6. The sub-millimeter set is n=16 across two scanner vendors (Siemens and GE), with no Philips or Toshiba coverage yet. "Equivalence" here means offline agreement of segmentation and radiomics on the reconstruction versus the lossless original.

It is not a prospective radiologist read, and production go-live needs that clinical validation. One of the five pre-registered gates passed as a statistical tie rather than a clean win: the 3D-versus-2D radiomics-ICC no-regression check missed by 0.0006 (3D 0.9920 versus 2D 0.9926), a margin that flips sign across cohorts with both models far above the 0.90 floor.

We read it as noise and document it as a tie, and a strict reading would not certify it. We validate with a single segmentation tool so far (TotalSegmentator), with nnU-Net as a planned independent cross-check. And on low-dose the neural model saturates around 53.5 dB, above the 48 dB diagnostic floor and cheaper than JPEG-2000 there, but a ceiling worth naming.

Frequently asked questions

Is lossy compression safe for diagnostic CT?

It can be, but only if you prove it per task rather than assume it from image quality. Our standard is that a downstream clinical task, radiomics stability or segmentation, reaches the same result on the reconstruction as on the lossless original, measured against a tool with a verified noise floor. Visual quality alone is not the bar.

What is BD-rate, and why is it not enough?

BD-rate summarizes how many fewer bits a codec spends at equal fidelity, usually against a proxy like PSNR. It is a bits metric, not a diagnosis metric. JPEG-2000 at 0.78 bpp keeps a fine BD-rate while collapsing a radiomics feature to ICC 0.49, so we gate on clinical tasks in addition to BD-rate, never instead of it.

Why does a 3D codec only help on thin slices?

Because inter-slice correlation rises as slices thin. On thick slices (2.5 mm) neighbors differ by about 41.7 HU on average, and coding across them saves little or costs bits. On sub-millimeter slices neighbors are nearly redundant, and a model that codes across slices captures that redundancy for a 12.5% bit-rate win. The benefit is gated to the thin end.

What is the takeaway for other neural medical-codec teams?

Measure the tool before you measure the codec. Use deterministic segmentation, full volumes, a self-agreement check, and only compare rates where the reference brackets your curve. Three separate times, one of those guards was the difference between a true verdict and a confident false one.