Validation of aortic valve calcification quantification on contrast-enhanced computed tomography against ex vivo gravimetric analysis: comparison of fixed Hounsfield unit thresholds and deep learning segmentation Authors: Daisong Jiang, Wanting Zhang, Lulu Liu, Geng Li, Xiaoke Shang, Nianguo Dong, Ye Yang, Jian Yang, Haibo Zhang, Qing Zhou, Chao Jian, Yuan Zhao, Buqing Ni, Yongfeng Shao, Jian Liu, Zhong Wu Published: 2026 ABSTRACT AbstractBackgroundAccurate quantification of aortic valve calcification (AVC) on contrast-enhanced computed tomography angiography (CTA) is pivotal for planning surgical and transcatheter aortic valve replacement. The optimal Hounsfield unit (HU) threshold for calcification detection on contrast-enhanced images remains unresolved, and every prior validation study has relied on non-contrast Agatston scoring—itself an imaging estimate—as the reference standard. This study validated two widely used fixed HU thresholds (450 HU and 850 HU) and a self-configuring nnU-Net deep learning model against ex vivo gravimetric calcium weight as an absolute physical ground truth.MethodsFour hundred patients were included in a retrospective cohort study with a pre-specified temporal validation split: 300 with CT-confirmed AVC and 100 with normal aortic valves. Fifty chronologically later AVC patients who underwent elective open surgical aortic valve replacement (SAVR) within seven days of clinically indicated pre-operative contrast-enhanced CTA formed the locked surgical validation cohort; their excised native leaflets underwent standardised high-temperature ashing (550 °C, 12 h) and analytical weighing (precision 0.1 mg) to obtain gravimetric calcium mass. The remaining 350 cases served exclusively for nnU-Net development (280 training / 70 internal validation). CT-derived calcium mass-equivalent estimates were quantified on the validation cohort and compared with gravimetric weight using Pearson and Spearman correlation and Bland-Altman analysis.ResultsThe nnU-Net achieved the strongest observed correlation with gravimetric weight (Pearson r = 0.967; bias + 6.2 mg; RMSE 13.7 mg), significantly outperforming the 450 HU threshold for correlation (r = 0.864; bias + 36.2 mg; RMSE 42.1 mg; Steiger p < 0.001) and showing a non-significant trend toward stronger correlation than 850 HU (r = 0.929; bias + 17.5 mg; RMSE 23.9 mg; Steiger p = 0.085). Compared with 850 HU, nnU-Net provided lower bias and RMSE, although the difference in Pearson r did not reach statistical significance. The 450 HU method exhibited significant proportional bias (p = 0.024), whereas neither 850 HU nor nnU-Net did. The nnU-Net achieved a mean Dice coefficient of 0.873 and intersection-over-union of 0.812.ConclusionsAgainst physically weighed calcium, nnU-Net deep learning segmentation provided the most favourable overall performance profile on contrast-enhanced CTA, with the lowest bias and RMSE and the strongest observed correlation. The improvement in Pearson correlation over 850 HU represented a non-significant trend, whereas the error and agreement metrics favoured nnU-Net. Among fixed thresholds, 850 HU substantially outperformed 450 HU, offering direct physical-rather than surrogate imaging-evidence to support 850 HU as the preferred fixed threshold in standard contrast-enhanced protocols.Keywords: Aortic valve calcification, Contrast-enhanced CT, Hounsfield unit threshold, nnU-Net, Deep learning, Gravimetric analysis, Surgical aortic valve replacement, Calcium quantification Background Degenerative calcific aortic stenosis is the most prevalent valvular heart disease in the industrialised world, affecting approximately 3% of adults over the age of 75, with its incidence projected to double as the global population ages [1]. Aortic valve calcification (AVC) is both the pathological driver of progressive stenosis and the primary imaging biomarker that guides intervention selection and technique. Before surgical aortic valve replacement (SAVR) or transcatheter aortic valve replacement (TAVR), quantification of AVC from computed tomography angiography (CTA) informs valve sizing, predicts the risk of paravalvular leak (PVL), identifies patients susceptible to annular injury, and supports decision-making in haemodynamically discordant cases [2, 3]. In non-contrast computed tomography (CT), the Agatston method provides a validated calcium score based on a fixed attenuation threshold of 130 HU [4]. For sex-specific severity classification, threshold values of ≥ 1,600 AU in women and ≥ 2,000 AU in men have been endorsed by contemporary guidelines [5]. Contrast-enhanced CTA—now the standard pre-procedural imaging modality—renders the 130 HU threshold wholly inapplicable, because intraluminal blood pool attenuation routinely reaches 250–450 HU and varies substantially with contrast dose, injection rate, cardiac output, and scanner acquisition parameters [6]. Consequently, a range of empirical fixed thresholds—most commonly 450 HU and 850 HU—have been adopted in clinical practice and research without consensus [7, 8]. The provenance of these two thresholds illustrates the ongoing ambiguity. Jilaihawi et al. demonstrated that, for non-contrast acquisitions, a threshold of 450 HU most discriminated patients at risk for significant PVL, whereas for contrast-enhanced scans a threshold of 850 HU was optimal, using PVL prediction as the clinical endpoint [9]. Angelillis et al. subsequently proposed that threshold selection should depend on the degree of left ventricular outflow tract (LVOT) luminal enhancement: 450 HU when LVOT attenuation falls below 300 HU, and 850 HU when it equals or exceeds 300 HU, each validated against non-contrast Agatston scores [10]. A systematic review encompassing 68 studies confirmed threshold values spanning 300–850 HU across the published literature and highlighted the absence of a universally accepted standard [11]. Crucially, every published validation study has used non-contrast Agatston scoring as its reference—an estimate that is itself subject to partial-volume averaging, scanner-dependent variability, and dependence on the empirical 130 HU detection threshold. No study has validated contrast-enhanced CTA calcium measurement against a true physical measurement of calcium mass. Deep learning-and specifically fully convolutional encoder-decoder networks-has transformed medical image segmentation over the past decade [12]. The nnU-Net framework of Isensee et al. represents a pivotal methodological advance in that it automatically configures its architecture, preprocessing pipeline, and training schedule based on dataset properties, consistently matching or outperforming hand-crafted architectures across diverse medical imaging tasks [13]. Recent cardiovascular CT studies have further shown the feasibility of automated or semi-automated aortic-valve and aortic-root segmentation and contrast-enhanced calcium quantification, although these methods have generally been validated against non-contrast CT, manual imaging labels, echocardiographic severity, or histology rather than against direct physical calcium mass [14–17]. Its application to cardiac calcification segmentation has not yet been validated against a gravimetric physical ground truth. The present study was designed to close this methodological gap. By enrolling patients undergoing open SAVR, we obtained surgically excised native aortic valve leaflets and quantified their calcium content by standardised high-temperature ashing and gravimetric weighing—yielding an absolute physical gold standard independent of any imaging modality. We then validated the 450 HU threshold, 850 HU threshold, and a trained nnU-Net model against this reference, providing what is, to our knowledge, the first direct physical validation of contrast-enhanced CTA calcium quantification in the aortic valve. Methods Study design and ethical approval This was a retrospective cohort study with a pre-specified temporal validation split. De-identified imaging, clinical, surgical, and laboratory data were collected from multiple participating centres. The temporal validation cohort was locked before model training and was not used for any model-development, hyperparameter-selection, or calibration-feedback step. The study was conducted in accordance with the Declaration of Helsinki; details of ethics approval and the waiver of individual informed consent for retrospective de-identified data are provided in the Declarations section. The overall study workflow is depicted in Fig. 1. Study design and patient flow diagram. A total of 400 patients were analysed in a retrospective cohort study with a temporal validation split. The locked surgical validation cohort comprised 50 chronologically later patients, from whom excised native aortic valve leaflets were processed by high-temperature ashing (550 °C, 12 h) to yield gravimetric calcium weight as the absolute gold standard; the model-development pool comprised 350 patients split into a training set (n = 280, 80%) and internal validation set (n = 70, 20%). The nnU-Net model and both fixed-threshold methods were evaluated exclusively on the locked surgical validation cohort. AVC: aortic valve calcification Patient population and eligibility Between January 2022 and December 2023, a consecutive series of 400 patients was enrolled across two groups. The AVC group comprised 300 patients with CT-confirmed aortic valve calcification identified from the cardiac CT databases of the participating centres. Eligible patients were aged 18 years or older, had undergone prospectively ECG-gated contrast-enhanced CTA of the aortic root within seven days of the index clinical evaluation, and had at least one calcified focus confirmed on axial CT images. Patients were excluded if they had undergone any prior aortic valve intervention, carried a congenital unicuspid valve diagnosis, or had CTA images of insufficient diagnostic quality for segmentation, defined as a motion-artefact score exceeding three on a validated five-point scale. The normal-valve group comprised 100 patients with no CT evidence of AVC, selected from the same database and matched approximately by age and sex to serve as negative-class training examples for the segmentation model. Using the temporal split, the 50 chronologically later eligible AVC patients who underwent elective open SAVR within seven days of clinically indicated index CTA formed the sealed surgical validation cohort. Contrast-enhanced CTA in these patients had been requested as part of routine heart-team evaluation of valve-intervention strategy and aortic-root anatomy, including patients initially assessed in a TAVR-versus-SAVR decision pathway and patients requiring detailed anatomic assessment before surgery. These patients ultimately underwent SAVR rather than TAVR after multidisciplinary review because of clinical or anatomic factors such as low-to-intermediate surgical risk, younger age, bicuspid morphology, concomitant surgical indications, coronary anatomy, or patient/surgeon preference. Their imaging data were kept under strict embargo throughout model development. The remaining 350 patients (250 with AVC and 100 with normal valves) constituted the model development pool, which was split 80/20 into training (n = 280) and internal validation (n = 70) subsets. Computed tomography acquisition All examinations were performed on dual-source CT scanners (Siemens SOMATOM Definition Flash or Force, Siemens Healthineers, Forchheim, Germany) at each participating centre using a harmonised prospectively ECG-triggered arterial-phase acquisition protocol. Images were acquired during inspiratory breath-hold at 70–80% of the R-R interval. Tube voltage was 100 kVp for all included scans; automated exposure control was used with a reference tube current-time product of 250 mAs. Detector collimation was 2 × 128 × 0.6 mm or 2 × 192 × 0.6 mm according to scanner generation, gantry rotation time was 250–280 ms, and pitch was adapted automatically to heart rate. Images were reconstructed with 0.75-mm slice thickness, 0.5-mm increment, and a medium-smooth B36f reconstruction kernel. The contrast agent (Iomeron 400, 400 mg iodine/mL) was injected at 5.0 mL/s via antecubital vein access using a weight-adapted volume of 60–80 mL followed by a 40-mL saline chaser. Bolus tracking was triggered in the ascending aorta at a threshold attenuation of 100 HU, followed by a 6-s scan delay. Luminal attenuation values in the ascending aorta and LVOT were measured by placing a circular region of interest (ROI) of at least 1 cm² at the centre of each vessel lumen on axial images at the level of the aortic annulus, avoiding wall structures. Ex vivo gravimetric calcium quantification Immediately upon surgical excision, each set of native aortic valve leaflets was placed in a pre-weighed platinum crucible, rinsed with physiological saline to remove blood and loose soft tissue, and transferred to a muffle furnace for ashing at 550 °C for 12 h. This temperature and duration have been established in the mineral analysis literature as sufficient to ensure complete combustion of organic material while preserving calcium-containing mineral phases, principally calcium hydroxyapatite [18]. Following ashing, crucibles were transferred to a sealed desiccator, allowed to equilibrate to ambient temperature, and weighed on a calibrated analytical microbalance (precision: 0.1 mg; Sartorius CPA26P). A blank crucible underwent identical thermal processing in parallel with each batch to correct for procedural weight gain from combustion by-products. The net ash mass after blank subtraction was recorded as the gravimetric calcium weight (mg) and served as the absolute gold standard for all subsequent comparisons. All gravimetric procedures were performed by a single trained laboratory technician who was blinded to all CT findings throughout. CT-based calcium quantification: fixed-threshold methods Post-processing was performed on a dedicated workstation using syngo.via software, version VB60. For both threshold methods, a three-dimensional ROI was defined manually on axial CT images to encompass only the native aortic valve leaflets-delimited inferiorly by the aortic annular plane and superiorly by the leaflet tip level-explicitly excluding annular, LVOT, and aortic wall calcification to match the anatomical scope of the excised specimens. The 450 HU method identified all voxels within the ROI with CT attenuation at or above 450 HU; the 850 HU method applied the same procedure at or above 850 HU. For all three methods, the binary mask defined the voxel set used for mass-equivalent estimation. Because segmentation masks are volumetric outputs, CT-derived calcium was converted to a protocol-specific calcium mass-equivalent (mCT, mg) using attenuation-weighted hydroxyapatite calibration rather than assuming the crystalline density of hydroxyapatite. A multi-insert hydroxyapatite calibration phantom with known concentrations was scanned with the same tube voltage and reconstruction kernel; linear regression of measured insert attenuation against known hydroxyapatite concentration generated CHA(HU) = alpha x HU + beta, where CHA is expressed in mg/cm³. For each method, mCT = sum[CHA(HUi) x Vvoxel/1000] over all voxels i in the segmentation mask, where Vvoxel is expressed in mm³ and 1000 converts mm³ to cm³; negative predicted concentrations were set to 0. The calibration equation was fixed before surgical validation and was not fitted to gravimetric results. Raw mask volume (mm³) was retained as the segmentation output, whereas the agreement metrics in Table 2 used mCT. All test-cohort cases were independently quantified by two cardiovascular radiologists (Reader 1: 12 years of experience; Reader 2: 8 years of experience) blinded to gravimetric results, and inter-reader reliability was assessed using the intraclass correlation coefficient (ICC; two-way mixed effects model, absolute agreement definition). Table 2Accuracy of CT-based calcium quantification methods against the gravimetric gold standard (n = 50)MethodPearson r (95% CI)Spearman ρBias (mg)95% LoA (mg)RMSE (mg)Proportional bias pvs. nnU-Net p†vs. 850 HU p†450 HU Threshold0.864 [0.764–0.922]0.927+ 36.2−6.6 to + 79.042.10.024< 0.0010.060850 HU Threshold0.929 [0.873–0.960]0.930+ 17.5−14.6 to + 49.623.90.1860.085—nnU-Net0.967 [0.940–0.981]0.968+ 6.2−18.1 to + 30.413.70.341——All Pearson correlations p < 0.001. †Steiger Z-test for dependent Pearson correlations (Z-statistics: nnU-Net vs. 450 HU, Z = 3.21; nnU-Net vs. 850 HU, Z = 1.72; 850 HU vs. 450 HU, Z = 1.88). Bias and LoA are expressed as CT-derived calcium mass-equivalent minus gravimetric calcium mass; positive values indicate imaging overestimation relative to the gravimetric gold standard, and negative values indicate imaging underestimation. LoA: 95% limits of agreement (bias ± 1.96 SD of differences). RMSE root mean square error. 95% CI for Pearson r by Fisher’s z-transformation Accuracy of CT-based calcium quantification methods against the gravimetric gold standard (n = 50) All Pearson correlations p < 0.001. †Steiger Z-test for dependent Pearson correlations (Z-statistics: nnU-Net vs. 450 HU, Z = 3.21; nnU-Net vs. 850 HU, Z = 1.72; 850 HU vs. 450 HU, Z = 1.88). Bias and LoA are expressed as CT-derived calcium mass-equivalent minus gravimetric calcium mass; positive values indicate imaging overestimation relative to the gravimetric gold standard, and negative values indicate imaging underestimation. LoA: 95% limits of agreement (bias ± 1.96 SD of differences). RMSE root mean square error. 95% CI for Pearson r by Fisher’s z-transformation nnU-Net model development Annotation of the 350-case development pool was performed semi-automatically by two board-certified cardiovascular radiologists (12 and 8 years of cardiovascular CT experience), both blinded to clinical outcomes, gravimetric results, and the validation-cohort measurements. Each case was first processed with the 850 HU threshold to generate an initial seed segmentation, after which both readers independently refined the masks slice-by-slice on axial images using multiplanar reformats when needed. Discrepancies were resolved by consensus in a joint reading session, and the resulting consensus masks defined the leaflet calcification labels for supervised training. Inter-annotator agreement was calculated before consensus. To assess intraobserver reproducibility, Reader 1 repeated mask correction in a random 50-case subset after a four-week washout interval, blinded to the original labels. Model development employed the nnU-Net framework (version 2) [13], which automatically determines the optimal network configuration (here: 3-D full-resolution), patch size, batch size, intensity normalisation, and data augmentation strategy from dataset fingerprint analysis. Training used the nnU-Net default composite loss function (sum of Dice loss and cross-entropy) with an SGD optimiser incorporating Nesterov momentum (initial learning rate 0.01; polynomial decay), run for 1,000 epochs with five-fold cross-validation on the development pool. Internal validation Dice scores were monitored across folds, and the final predictions were generated by ensemble averaging of all five fold models. Three independent training runs with fixed but different random seeds were completed; all performance metrics are reported as mean ± standard deviation across runs. Training and inference were performed on a workstation equipped with an NVIDIA RTX 3090 GPU (24 GB VRAM) and 128 GB RAM. Post-processing retained only connected components larger than 1 mm³ and applied morphological closing (3 × 3 kernel) to smooth boundaries. Test-set inference was performed once after model finalisation, with no iterative tuning based on test-set feedback. Mean inference time was 24 ± 6 s per case for mask generation and less than 90 s per case including preprocessing, post-processing, and mass-equivalent calculation. Statistical analysis Statistical analysis was performed using Python 3.11 (SciPy 1.11; pandas 2.0; pingouin 0.5). Data are reported as mean ± standard deviation for normally distributed continuous variables and as median [interquartile range] for skewed distributions, following Shapiro-Wilk testing. Pearson r was used when both variables satisfied normality; Spearman’s rho was computed as a non-parametric complement in all cases. Agreement between CT-derived calcium mass-equivalent (mCT, mg) and gravimetric weight was evaluated by Bland-Altman analysis, reporting bias (mean difference, mCT minus gravimetric mass), 95% LoA (bias ± 1.96 × SD of differences), and RMSE. A positive bias therefore denotes imaging overestimation relative to gravimetric calcium mass, whereas a negative value denotes imaging underestimation. Proportional bias was assessed by regressing the difference score against the mean of the two methods; significance of the regression slope was tested by t-test. Pairwise differences in correlation coefficients were evaluated using the Steiger Z-test for dependent samples [19]. Segmentation model performance was reported as Dice coefficient, IoU, sensitivity, specificity, and precision. Inter-reader, inter-annotator, and intraobserver reproducibility were quantified by Dice coefficient and/or ICC as appropriate, with 95% confidence intervals estimated by bootstrapping (1,000 iterations). Two-tailed p < 0.05 was considered statistically significant. Results Patient characteristics Baseline demographic and clinical characteristics are presented in Table 1. The surgical test cohort (n = 50) and the model development pool (n = 350) were well balanced in age (70.1 ± 8.7 vs. 68.4 ± 9.2 years; p = 0.27), sex distribution (58.0% vs. 54.0% male; p = 0.59), body mass index, and all major comorbidities (all p > 0.20). Bicuspid aortic valve morphology was identified in 8 test-cohort patients (16.0%) and 52 development-pool patients (14.9%; p = 0.83). In the test cohort, mean LVOT attenuation was 318.4 ± 74.2 HU (range: 187–486 HU) and mean ascending aortic attenuation was 412.6 ± 68.3 HU. Twenty-three patients (46.0%) had LVOT attenuation below the pre-specified 300 HU stratification threshold, and the remaining 27 (54.0%) had LVOT attenuation at or above 300 HU. The median gravimetric calcium weight of the excised leaflets was 111.0 mg (IQR: 73.4–146.3 mg). Table 1Baseline demographic and clinical characteristicsCharacteristicDevelopment Pool (n = 350)Surgical Test Cohort (n = 50)p-valueDemographicsAge (years), mean ± SD68.4 ± 9.270.1 ± 8.70.27Male sex, n (%)189 (54.0%)29 (58.0%)0.59Body mass index (kg/m²), mean ± SD24.8 ± 3.425.1 ± 3.10.60ComorbiditiesHypertension, n (%)214 (61.1%)32 (64.0%)0.70Diabetes mellitus, n (%)87 (24.9%)14 (28.0%)0.63Current or ex-smoker, n (%)126 (36.0%)18 (36.0%)0.99Coronary artery disease, n (%)98 (28.0%)15 (30.0%)0.78eGFR (mL/min/1.73 m²), mean ± SD72.3 ± 18.669.8 ± 17.40.41Aortic Valve MorphologyBicuspid aortic valve, n (%)52 (14.9%)8 (16.0%)0.83Tricuspid aortic valve, n (%)298 (85.1%)42 (84.0%)0.83CT Acquisition — Test Cohort OnlyLVOT attenuation (HU), mean ± SD—318.4 ± 74.2—Ascending aorta attenuation (HU), mean ± SD—412.6 ± 68.3—LVOT attenuation < 300 HU, n (%)—23 (46.0%)—Gravimetric Gold Standard — Test Cohort OnlyCalcium weight (mg), median [IQR]—111.0 [73.4–146.3]—eGFR estimated glomerular filtration rate, LVOT left ventricular outflow tract, IQR interquartile range. p-values compare the surgical test cohort against the development pool using Student’s t-test or chi-squared test as appropriate Baseline demographic and clinical characteristics eGFR estimated glomerular filtration rate, LVOT left ventricular outflow tract, IQR interquartile range. p-values compare the surgical test cohort against the development pool using Student’s t-test or chi-squared test as appropriate Agreement of CT methods with gravimetric gold standard Pearson and Spearman correlation statistics and Bland-Altman agreement parameters for all three methods are summarised in Table 2; corresponding scatter plots and Bland-Altman plots are presented in Figs. 2 and 3, respectively. All three CT methods demonstrated significant positive correlation with gravimetric calcium weight. Scatter plots illustrating the correlation between CT-derived calcium mass-equivalent (mg) and ex vivo gravimetric calcium weight (mg) in the held-out surgical validation cohort (n = 50). Panel A: 450 HU threshold (Pearson r = 0.864). Panel B: 850 HU threshold (r = 0.929). Panel C: nnU-Net segmentation (r = 0.967). Solid lines represent linear regression fits; dashed lines indicate the line of identity. All correlations p < 0.001 Bland-Altman plots depicting agreement between each CT-derived calcium mass-equivalent estimate and the gravimetric gold standard (n = 50). Panel A: 450 HU threshold (bias + 36.2 mg; LoA − 6.6 to + 79.0 mg). Panel B: 850 HU threshold (bias + 17.5 mg; LoA − 14.6 to + 49.6 mg). Panel C: nnU-Net segmentation (bias + 6.2 mg; LoA − 18.1 to + 30.4 mg). Solid lines indicate mean bias; dashed lines indicate the 95% limits of agreement. A positive bias denotes CT overestimation relative to gravimetric mass, and a negative value denotes CT underestimation. The nnU-Net exhibits the smallest bias and narrowest limits of agreement. Note the dispersion pattern in Panel A indicating proportional overestimation by the 450 HU method at higher calcium loads The nnU-Net achieved the strongest observed linear correlation (Pearson r = 0.967, 95% CI: 0.940–0.981; Spearman rho = 0.958; both p < 0.001). Its advantage over the 450 HU threshold was statistically significant (Steiger Z = 3.21; p < 0.001), whereas its higher Pearson r compared with the 850 HU threshold represented a non-significant trend (Steiger Z = 1.72; p = 0.085) and should not be interpreted as statistically superior correlation at the conventional p < 0.05 threshold. The 850 HU threshold (r = 0.929, 95% CI: 0.873–0.960) and the 450 HU threshold (r = 0.864, 95% CI: 0.764–0.922) did not differ significantly in Pearson r (Steiger Z = 1.88; p = 0.060), although the point estimates and all Bland-Altman metrics favoured 850 HU. Bland-Altman analysis revealed positive systematic bias for all three methods, indicating a consistent tendency to overestimate calcium burden relative to the physical gold standard (Fig. 3). The 450 HU method produced the largest bias (+ 36.2 mg) and the widest LoA (− 6.6 to + 79.0 mg; RMSE = 42.1 mg). Regression of the difference score against the mean was statistically significant for the 450 HU method (slope = 0.31; p = 0.024), indicating proportional bias—overestimation increased with greater calcium load. The 850 HU threshold achieved a substantially smaller bias (+ 17.5 mg; LoA: −14.6 to + 49.6 mg; RMSE = 23.9 mg) and showed no significant proportional bias (slope = 0.12; p = 0.186). The nnU-Net exhibited the smallest bias of the three methods (+ 6.2 mg; LoA: −18.1 to + 30.4 mg; RMSE = 13.7 mg) and no significant proportional bias (slope = 0.08; p = 0.341). Inter-reader ICC for the fixed-threshold methods was excellent: 0.923 (95% CI: 0.888–0.948) for 450 HU and 0.941 (95% CI: 0.912–0.962) for 850 HU. nnU-Net segmentation performance On the held-out surgical validation cohort, the nnU-Net achieved a mean Dice coefficient of 0.873 ± 0.021, IoU of 0.812 ± 0.025, sensitivity of 0.901 ± 0.018, specificity of 0.968 ± 0.012, and precision of 0.886 ± 0.023 (Table 3; Fig. 4). Inter-annotator agreement for the development-set reference labels was high (mean Dice 0.891 ± 0.019; ICC 0.941, 95% CI: 0.912–0.962), and intraobserver repeatability in the 50-case repeat-annotation subset was also excellent (mean Dice 0.908 ± 0.017; ICC 0.956, 95% CI: 0.932–0.972), indicating that the model was trained on consistent, reproducible annotations. All five segmentation metrics exceeded or closely approached the pre-specified clinical acceptance threshold of 0.85, with IoU (0.812) representing the only metric that did not meet this threshold. Table 3nnU-Net segmentation performance on the held-out surgical validation Set (n = 50)MetricMean ± SD95% CIDice Coefficient0.873 ± 0.021[0.864–0.882]Intersection over Union (IoU)0.812 ± 0.025[0.805–0.820]Sensitivity (Recall)0.901 ± 0.018[0.895–0.908]Specificity0.968 ± 0.012[0.963–0.973]Precision0.886 ± 0.023[0.879–0.893]Inter-annotator Dice (reference labels)0.891 ± 0.019[0.884–0.899]Inter-reader ICC (threshold quantification)0.941[0.912–0.962]Inter-annotator ICC0.941[0.912–0.962]Intra-annotator Dice0.908 ± 0.017[0.901–0.915]Intra-annotator ICC0.956[0.932–0.972]Values are mean ± SD across three independent training runs with different random seeds unless otherwise specified. 95% CI estimated by bootstrapping (1,000 iterations) on the per-patient test-set predictions. Inter-annotator and intra-annotator Dice/ICC metrics refer to reference-label mask reproducibility. Inter-annotator metrics were calculated before consensus mask generation; intra-annotator metrics were calculated in a random 50-case repeat-annotation subset after a four-week washout interval. ICC intraclass correlation coefficient (two-way mixed effects, absolute agreement) nnU-Net segmentation performance on the held-out surgical validation Set (n = 50) Values are mean ± SD across three independent training runs with different random seeds unless otherwise specified. 95% CI estimated by bootstrapping (1,000 iterations) on the per-patient test-set predictions. Inter-annotator and intra-annotator Dice/ICC metrics refer to reference-label mask reproducibility. Inter-annotator metrics were calculated before consensus mask generation; intra-annotator metrics were calculated in a random 50-case repeat-annotation subset after a four-week washout interval. ICC intraclass correlation coefficient (two-way mixed effects, absolute agreement) Bar chart of nnU-Net segmentation performance metrics on the held-out surgical test cohort (n = 50): Dice coefficient (0.873), IoU (0.812), sensitivity (0.901), specificity (0.968), and precision (0.886). Bars represent mean values across three independent training runs; error bars indicate ± standard deviation. Individual training-run values are overlaid as black dots. The dashed line indicates a clinical acceptance threshold of 0.85 Subgroup analysis by LVOT attenuation Among the 23 patients with LVOT attenuation below 300 HU, the 450 HU and 850 HU thresholds performed comparably (r = 0.901 vs. r = 0.912; Steiger p = 0.81), consistent with prior imaging-based evidence [10]. In the 27 patients with LVOT attenuation at or above 300 HU, the 850 HU threshold significantly outperformed the 450 HU threshold (r = 0.972 vs. r = 0.889; Steiger p = 0.018), while the nnU-Net maintained the highest correlation in both subgroups (LVOT < 300 HU: r = 0.966; LVOT ≥ 300 HU: r = 0.983), supporting its robustness to variability in luminal contrast attenuation. Discussion This study provides the first validation of contrast-enhanced CTA AVC quantification against a true physical reference—ex vivo gravimetric calcium mass from surgically excised leaflets. Three principal findings are reported. The first finding is that the nnU-Net deep learning model provided the most favourable overall performance profile for AVC quantification. With a bias of only + 6.2 mg and an RMSE of 13.7 mg, nnU-Net showed lower error than both threshold-based methods. It also had the strongest observed Pearson correlation with gravimetric weight; however, the correlation advantage was statistically significant versus 450 HU but not versus 850 HU. Therefore, the present data support nnU-Net primarily on the basis of improved agreement and error metrics relative to 850 HU, rather than a statistically proven superiority in Pearson correlation. This performance pattern is conceptually explicable: fixed HU thresholds operate purely on local voxel intensity and cannot distinguish calcified tissue from adjacent high-density structures-such as contrast-enhanced blood pool, annular fibrous tissue, or aortic wall-particularly when luminal attenuation is variable. By contrast, nnU-Net leverages spatial context, three-dimensional structural continuity, and learned tissue-boundary features to resolve ambiguous regions. The nnU-Net framework is particularly well suited to this application because its self-configuring design accounts for the specific voxel spacing, intensity distribution, and class imbalance inherent to CTA-based leaflet calcium segmentation without requiring manual architecture engineering [13, 14]. The second finding is that, among fixed-threshold methods, 850 HU substantially outperforms 450 HU under standard contrast-enhanced CTA conditions. The 450 HU method exhibited a bias of + 36.2 mg with an RMSE of 42.1 mg—more than three times the RMSE of the nnU-Net—and demonstrated statistically significant proportional bias, meaning that its overestimation escalated progressively with increasing calcium load. This pattern has important clinical implications: patients with severe calcification, who are precisely the highest-risk group for PVL and annular injury, are also those in whom the 450 HU threshold is least accurate. The 850 HU threshold was free from proportional bias and achieved a bias of + 17.5 mg, confirming that it is the more appropriate fixed-threshold choice for standard contrast protocols. These results are consistent with the clinical observations of Jilaihawi et al. [9] and the imaging-based findings of Angelillis et al. [10], but elevate the evidence to a physically grounded standard—an advance over all prior validation methodologies. The third finding concerns the influence of luminal contrast attenuation and scanner-dependent variability. The subgroup analysis confirms that when LVOT enhancement is modest (< 300 HU), the performance gap between 450 HU and 850 HU narrows to non-significance, in keeping with the conditional recommendation of Angelillis et al. [10]. In patients with higher LVOT enhancement (≥ 300 HU), the 850 HU threshold was significantly more accurate than 450 HU, while the nnU-Net remained numerically strongest in both subgroups. Fixed thresholds are inherently sensitive to contrast concentration, kVp, reconstruction kernel, image noise, and partial-volume effects because all of these factors shift voxel attenuation distributions. Deep learning segmentation may be less sensitive to a single HU cut-off because it integrates anatomical context and shape information, but it is not immune to domain shift; vendor, kernel, tube-voltage, and contrast-protocol differences can still alter image appearance and therefore require external validation and, when necessary, site-specific recalibration or domain adaptation. Several limitations of this study warrant consideration. First, the surgical validation cohort was a selected subgroup of patients who underwent clinically indicated contrast-enhanced CTA before ultimately receiving SAVR after heart-team review. Some patients were initially assessed in a TAVR-versus-SAVR decision pathway, and others required detailed anatomic assessment before surgery. This may enrich the validation cohort for patients with more complex anatomy, bicuspid morphology, severe calcification, or concomitant surgical indications, and the results may not fully represent an unselected AS population or patients undergoing routine SAVR without CT. Second, although data were collected from multiple centres, all sites used the same scanner vendor platform and a harmonised acquisition protocol; generalisability to other vendors and protocols requires further validation. nnU-Net performance, phantom calibration, and optimal threshold behaviour may differ across scanner vendors, reconstruction kernels, tube voltage protocols, iterative reconstruction settings, or contrast injection schemes, and external multi-centre validation is essential before clinical adoption. Third, gravimetric ashing yields total mineral mass without distinguishing calcium hydroxyapatite from other calcium salts; however, hydroxyapatite constitutes the dominant mineral phase in valvular calcification [18], and systematic compositional differences between patients are unlikely to introduce differential bias. Fourth, CT-derived calcium mass-equivalent estimates depend on the calibration equation and should not be interpreted as direct chemical calcium mass independent of acquisition protocol. Fifth, the surgical validation cohort comprised 50 patients-sufficient for correlation and agreement analysis but limiting the precision of subgroup estimates; larger test populations would strengthen subgroup inferences, particularly for bicuspid morphology. Sixth, the nnU-Net model relied on five-fold cross-validation for internal model selection, which introduces the possibility of minor overfitting to the development distribution. Finally, dynamic threshold strategies-such as the luminal attenuation + 100 HU protocol [7], false-positive-rate approaches [20], and the EuroIntervention correction formula [21]-were not compared against the gravimetric standard, representing a natural extension of this work. The clinical implications are direct. Centres performing pre-procedural CTA without a paired non-contrast sequence should prefer the 850 HU threshold over 450 HU, and should use 450 HU only when LVOT attenuation is demonstrably below 300 HU. Where deep learning infrastructure is available, nnU-Net segmentation may further reduce systematic overestimation and provide attenuation-robust leaflet-level calcium mapping. In TAVR planning, such maps could complement standard Agatston scoring by localising bulky leaflet calcification, commissural or raphe-related calcium, and eccentric calcium distributions that are relevant to PVL risk, annular or root injury, need for balloon pre- or post-dilatation, valve sizing strategy, and procedural risk communication [2, 3, 11, 15–17]. The computational burden appears clinically feasible: in this implementation, inference and post-processing required less than 90 s per case on a single workstation GPU, allowing integration as an offline pre-procedural post-processing step. The trained model and inference code will be made publicly available at [repository URL] to facilitate external validation and integration into clinical workflows. Conclusions Using ex vivo gravimetric analysis of surgically excised aortic valve leaflets as an absolute physical gold standard, nnU-Net deep learning segmentation provided the lowest bias and RMSE for AVC quantification on contrast-enhanced CTA (r = 0.967; bias + 6.2 mg; RMSE 13.7 mg). Its Pearson correlation was significantly higher than the 450 HU fixed threshold (p < 0.001), while its higher correlation than 850 HU represented a non-significant trend (p = 0.085). Among threshold methods, 850 HU was substantially more accurate than 450 HU (bias + 17.5 mg vs. +36.2 mg; RMSE 23.9 mg vs. 42.1 mg), and the 450 HU method additionally exhibited significant proportional bias that escalated with calcium burden. These findings provide physically grounded evidence to inform HU threshold selection for contrast-enhanced CTA AVC quantification and support further external validation of nnU-Net segmentation for pre-operative assessment of aortic valve disease. Acknowledgements Not applicable. Abbreviations Agatston units Aortic valve calcification Confidence interval Computed tomography Computed tomography angiography Electrocardiogram Estimated glomerular filtration rate Hounsfield unit Intraclass correlation coefficient Intersection over union Interquartile range Limits of agreement Left ventricular outflow tract no-new-Net (self-configuring deep learning framework) Paravalvular leak Root mean square error Region of interest Surgical aortic valve replacement Standard deviation Stochastic gradient descent Transcatheter aortic valve replacement Authors’ contributions DSJ and WTZ contributed equally to this work as co-first authors. DSJ conceived the study, developed the nnU-Net model, performed the statistical analysis, and drafted the manuscript. WTZ designed the study, performed CT-based calcium quantification, coordinated multi-centre data collection, and drafted the manuscript. LLL contributed to image annotation, data curation, and critical revision of the manuscript. GL and XKS performed surgical specimen collection and contributed to data acquisition. NGD supervised the surgical procedures and contributed to study design. YY, JY, HBZ, QZ, CJ, and YZ contributed to patient recruitment, data acquisition, and image quality assessment at their respective centres. BQN contributed to gravimetric laboratory procedures and data interpretation. YFS contributed to study coordination and critical revision of the manuscript. JL and ZW are co-corresponding authors; both supervised the project, secured funding, and critically revised the manuscript for important intellectual content. All authors read and approved the final manuscript. Data availability No datasets were generated or analysed during the current study. Declarations Ethics approval and consent to participate This study was approved by the Ethics Committee of West China Hospital (Approval No.: HX-2022-17), which served as the lead ethics review authority for all participating sites. The institutional review board approved this retrospective cohort study and waived the requirement for individual informed consent because all analyses used de-identified imaging, clinical, surgical, and laboratory data. Surgically excised native leaflets were clinically discarded specimens obtained during standard SAVR and were processed under the IRB-approved retrospective protocol. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Jian Liu, Email: jameslau1984@sina.com. Zhong Wu, Email: wuzhong71@scu.edu.cn. References Associated Data Data Availability Statement No datasets were generated or analysed during the current study.