Most kidney-tumor detectors are trained on contrast-enhanced CT scans of patients already headed to surgery. Screening a high-risk population such as long-term dialysis patients looks different. These scans are taken without intravenous contrast, the tumors are smaller, and the model may have to run in low precision on cheap hardware. The deployed model therefore faces two stacked perturbations: its inputs lose soft-tissue contrast, and its weights are rounded to a coarse grid. This proposal tests whether one property of training, landing in a flat minimum (a region of parameter space where loss changes little when weights move), protects small-lesion sensitivity against both perturbations at once, and whether that protection holds beyond what intensity augmentation already provides.
Measuring this honestly is harder than training it. In hospital archives, unenhanced scans differ from contrast scans in who was imaged and why. They come from different patients, tumor sizes, scanners and indications, so a naive comparison mixes "no contrast" with "different patients." We borrow the logic of the retrospective cohort study to separate these. The cleanest estimate comes from multiphase renal CT exams, in which the same patient is scanned with and without contrast in one session, so case mix is held fixed. For transferring results to an unpaired screening cohort, we reweight cases so that the source population matches the screening population's case mix.
Core Claim
The claim has three parts. First, a 3D segmentation model trained toward flat minima, using the functional variant of sharpness-aware minimization, will lose less lesion-level sensitivity on small renal tumors when moved from contrast to unenhanced CT. Second, this benefit remains when compared against a strong augmentation baseline given equal compute. Third, the same model will degrade less under 4-bit weight quantization, and a sharpness measurement taken only on contrast-enhanced training data will predict, across checkpoints, how much sensitivity survives the combined shift.
The claim is falsifiable at each step. If augmentation alone closes the contrast gap as well as flatness does, the input-shift half of the idea fails, and we will say so.
Background and Positioning
The screening motivation comes from Ishikawa's nationwide Japanese dialysis surveys. In Renal cell carcinoma in chronic hemodialysis patients in Japan, dialysis patients carried a 12- to 40-fold elevated renal cell carcinoma risk, and acquired cystic kidney disease was present in 80% of cases. Only 17.4% of the 1988 cases were symptomatic. Later survey rounds (1998, 2004) report that 88–95% of tumors were found by imaging surveillance while still asymptomatic. Those studies were purely epidemiological and involved no automated detection. They define our target population: cyst-laden kidneys, small tumors (localized mean 4.6 cm versus 8.8 cm for metastatic cases), and scans that avoid contrast.
On the imaging side, the KiTS19 challenge built its benchmark from 300 contrast-enhanced scans of surgical patients. Its winning 3D U-Net reached tumor Dice 0.851, against about 0.580 for standard baselines. This source cohort is exactly the one that misrepresents screening. The review by McGough et al. identifies the contrast-to-unenhanced drop as a central obstacle. It reports that radiomic analysis on non-contrast CT reached AUC 0.85 for benign-versus-malignant classification, but it proposes no training remedy.
Three lines of work supply the method.
- Compression is safe at 8 bits. Qu et al. show that 8-bit post-training quantization of 3D medical segmenters on TensorRT costs less than 0.005 mean Dice, but they never touch curvature or go below 8 bits.
- Flatness helps compression and shift in other domains. Na et al. show that sharpness-aware training makes language models more compressible. Jiang et al. (FQAT) show that ordinary quantization-aware training sharpens the loss surface and hurts out-of-distribution accuracy, and that flatness-aware quantization-aware training recovers it (+5.24% out-of-distribution accuracy on PACS at 3 bits). Their benchmarks are photos versus sketches, not a physical change in how images are acquired.
- Plain sharpness-aware training can mislead. Avoiding spurious sharpness minimization shows that standard sharpness-aware minimization can "flatten" a model by inflating logit scale rather than smoothing the input-to-output function. The paper proposes a functional variant that discards the logit path. This matters for us because screening triage runs on fixed score thresholds, and logit inflation would make those thresholds transfer badly across domains.
Two papers constrain the design. Catalan-Tatjer et al. find that longer training drives models into sharper, quantization-fragile minima, so we must control for training length. Sakata et al. show that at very low bit widths a badly chosen quantization range fragments the loss surface into many poor local minima. That result warns us that flatness at full precision may not survive 2-bit rounding, so we stop at 4 bits.
Research Program
Paired cohort. We will assemble a retrospective institutional cohort of multiphase renal-mass CT exams in which an unenhanced phase and a contrast phase were acquired in the same session. Tumors are annotated on the contrast phase, and the masks are carried to the unenhanced phase by registration within the exam. A radiologist checks the registered masks on a subset.
Screening-like cohort. We will assemble a second, unpaired cohort of unenhanced abdominal CT from chronic dialysis patients. Labels come from subsequent pathology or follow-up imaging. This cohort mirrors Ishikawa's population.
Case-mix adjustment. For the unpaired cohort, we fit a propensity model: a classifier that predicts, from covariates, whether a lesion comes from the screening cohort or the source cohort. The covariates are tumor size, cyst burden, age, dialysis duration, scanner vendor and slice thickness. Source-cohort test lesions are reweighted by the resulting odds. This lets us report what the source model's sensitivity would be on the screening case mix, separately from the effect of contrast.
Training arms. All models are nnU-Net-style 3D U-Nets trained on KiTS19/KiTS21-style contrast data. The arms differ only in the optimizer objective:
- standard training;
- sharpness-aware minimization;
- its preconditioned functional variant;
- training with random Gaussian weight noise, a cheap stand-in for flatness.
Each model is then compressed four ways: 8-bit post-training quantization (following Qu et al.), 4-bit post-training quantization, 4-bit quantization-aware training, and FQAT.
Sharpness measurement. For every checkpoint we record, on contrast training data only, two quantities. The first is the worst-case loss increase within a small weight radius. The second is the top Hessian eigenvalue, measured through the network function rather than logit scale. We then test whether these source-only numbers predict retained sensitivity under the stacked shift.
Baselines
Naive reference. Standard nnU-Net training on contrast data, with default augmentation and 8-bit post-training quantization, evaluated directly on unenhanced scans. This is what a hospital would deploy today. Qu et al. report under 0.005 mean Dice loss from 8-bit quantization, so this baseline isolates the contrast gap. We found no result on Lacuna for kidney-tumor segmentation or detection on unenhanced CT, so we will establish that number ourselves.
Strongest input-shift competitor. The same network trained with aggressive contrast-simulating intensity augmentation: randomized attenuation of enhancing tissue, histogram remapping toward unenhanced statistics, and simulated low-dose noise. The strongest in-domain reference point is the KiTS19 winner's tumor Dice of 0.851 on contrast scans. It is reported only in-domain and serves as a ceiling, not a target. To keep the comparison fair, the augmentation arm receives twice the training steps, matching the doubled cost of sharpness-aware updates, and the same hyperparameter search budget over augmentation strength.
Strongest compression competitor. FQAT, run as described by Jiang et al. at 4 bits on the same data. The only number the file gives is its PACS 3-bit out-of-distribution gain of +5.24%; no result was found for 3D medical models.
Ablations isolating flatness. Three ablations separate flatness from its side effects. Random-weight-noise training separates generic regularization from worst-case flatness. Plain versus functional sharpness-aware minimization separates genuine smoothing from logit inflation. Training-length-matched checkpoints separate flatness from the effect Catalan-Tatjer et al. describe.
What counts as winning. The proposal beats its baselines if the flatness arm shows higher small-lesion (<4 cm) sensitivity at a fixed false-positive rate per scan on the paired unenhanced phase than the matched-compute augmentation arm. The confidence interval comes from a paired bootstrap over patients. The flatness arm must also show a smaller 8-to-4-bit sensitivity drop than 4-bit quantization-aware training alone.
Why not just…?
Why not just use contrast-simulating augmentation? This is the real threat. Augmentation directly targets the known physical change, while flatness targets it only indirectly. No paper in the file compares the two on medical contrast shift. The matched-compute comparison on paired exams settles it. We also run augmentation combined with flatness. If the combination adds nothing over augmentation alone, we drop the input-shift claim and keep only the compression result.
Why not just fine-tune on unenhanced scans with pseudo-labels? Pseudo-labeling is powerful. In A-Eval it raised FLARE22 mean Dice from 82.37% to 87.91%. But it needs a sizable unlabeled target archive per site, and its pseudo-labels are least reliable on exactly the small, low-contrast tumors screening depends on. We include a pseudo-label fine-tuning arm and report whether flatness still helps as its initialization.
Why not just stay at 8 bits? Qu et al. show 8 bits is nearly free, so if the target hardware supports it, the compression half of this proposal is unnecessary. The 4-bit arm is framed as a stress test: flatness should matter where quantization noise is large. We report 8-bit results so readers can judge whether 4 bits is worth pursuing.
Why not just compare accuracy across archives without adjustment? Because screening tumors are smaller and sit in cystic kidneys, a naive comparison will overstate the contrast gap. We compute the naive unpaired estimate alongside the paired and reweighted estimates and report the discrepancy directly.
Decisive Checks
The first decisive check is the paired, matched-compute comparison of flatness against augmentation on small-lesion sensitivity. Everything else depends on it.
The second check is whether source-only sharpness, especially the functional measure, rank-correlates with retained sensitivity under the combined contrast-plus-4-bit shift across checkpoints and arms. The correlation is reported with and without controlling for training length.
The third check applies the winning model to the dialysis cohort. We report reweighted sensitivity, the fraction of scans a radiologist could skip at 95% sensitivity, and the drift of the triage threshold between domains. The threshold-drift measure is where functional and plain sharpness-aware minimization should differ most.
A negative result on the first check, alongside positive results on the second and third, would still be useful. It would mean flatness is a compression tool rather than a domain-shift tool in this setting.
References
- Renal cell carcinoma in chronic hemodialysis patients in Japan
- Renal cell carcinoma in chronic hemodialysis patients-A 1998 questionnaire study.
- Present status of renal cell carcinoma in dialysis patients: A questionnaire study in 2004 and review of past questionnaires since 1982
- The state of the art in kidney and kidney tumor segmentation in contrast-enhanced CT imaging: Results of the KiTS19 Challenge
- Artificial intelligence for early detection of renal cancer in computed tomography: A review
- Post-Training Quantization for 3D Medical Image Segmentation: A Practical Study on Real Inference Engines
- Train Flat, Then Compress: Sharpness-Aware Minimization Learns More Compressible Models
- Quantization Meets OOD: Generalizable Quantization-aware Training from a Flatness Perspective
- Avoiding spurious sharpness minimization broadens applicability of SAM
- Training Dynamics Impact Post-Training Quantization Robustness
- Effect of Weight Quantization on Learning Models by Typical Case Analysis
- A-Eval: A Benchmark for Cross-Dataset Evaluation of Abdominal Multi-Organ Segmentation