An AI method for segmenting the whole heart recorded the highest reported mean Dice scores on held-out CT and MRI acquisition sites in an arXiv preprint. Dice is the study's main overlap score for predicted cardiac structures and their annotations. On CT, the proposed method scored 0.9135 at held-out site G, versus 0.8343 for the strongest non-final baseline. On MRI, it scored 0.7830 at held-out site E, compared with 0.7790 for nnU-Netv2, making the MRI difference much smaller.
The question was whether a modality-routed, site-motivated appearance-augmentation pipeline could improve whole-heart segmentation generalization from limited CT and MRI training data to unseen acquisition distributions. The study evaluates that question with held-out-site comparisons.
The test followed acquisition sites
The data covered CT sites A, B and G, plus MRI site groups C/D and E. MRI C and D were pooled because separable site labels were unavailable. The reported cases were 20 at each CT site, 20 in the MRI C/D group and 26 at MRI site E, for 106 cases in all.
During development, CT models trained on A+B and were tested on G, while MRI models trained on C/D and were tested on E. The study also describes retraining final submissions on all available labeled cases and comparing Dice and HD95 with named baselines under the same held-out-site protocol.
The recipe changed appearance, not geometry
The pipeline first routed each scan to a CT or MRI branch using a deterministic intensity-based rule. Each branch used a TotalSegmentator-initialized nnU-Net model to produce per-voxel class probabilities. At inference, a class-wise largest-connected-component step kept the largest connected predicted region for each class.
Training combined two label-preserving appearance changes. Bias Field supplied a smooth spatial intensity perturbation, while Bezier applied a nonlinear intensity remapping. The retained recipe combined Bias Field + Bezier augmentation with the same class-wise largest-connected-component cleanup.
The augmented-distribution analysis reported expanded intensity and frequency variability after Bias + Bezier while preserving geometric spacing.
CT showed the clearest numerical gap
The CT ablation gave the highest listed score. Mean Dice was 0.8350 for Base, 0.8586 for Base + LCC, 0.8949 for Bias, 0.8876 for Bezier, 0.9036 for Bias + Bezier and 0.9135 for Bias + Bezier + LCC. The full recipe's 0.9135 also matches the proposed method's held-out CT result.
The MRI scores were closer together. Mean Dice was 0.7695 for Base, 0.7715 for Base + LCC, 0.7782 for Bezier, 0.7740 for Bias, 0.7807 for Bias + Bezier and 0.7830 for Bias + Bezier + LCC. The full recipe was the best listed MRI configuration, while nnU-Netv2 scored 0.7790.
Boundary performance was also summarized with HD95, the 95th-percentile Hausdorff distance. In the ablations, CT HD95 was 33.25 mm for Base, 9.12 mm for Base + LCC and 5.22 mm for Bias + Bezier + LCC. MRI HD95 was 24.50, 20.13 and 19.30 mm for those same variants.
The result remains tied to the tested sites
Numerically, the case for the recipe is stronger on CT than MRI: the CT comparison shows a wider gap, while the MRI edge over nnU-Netv2 is small. The reported results support a possible cross-site robustness benefit in this segmentation test, but they do not by themselves extend beyond the tested setup.
The study's main caveat is how narrowly its site analysis can be read. The authors state that the site-characterization analysis is specific to the available CARE-WHS training data. The development evaluation used one held-out site for each modality, CT site G and MRI site E, and MRI C and D could not be separated into individual sites. These details limit how broadly the reported site patterns can be generalized.
Publication and disclosures
The document is an arXiv preprint. It reports support from NSF grants IIS-2434967 and CNS-2431725, the NAIRR Pilot and TACC, the Computing Research Association, Purdue ACC Curricular Innovation Seed Grant and Purdue Applied AI Research Center. The authors report no competing interests relevant to the article.
Paper data and sources
Original title: Improving Cross-Site Whole-Heart Segmentation
Authors: Tanish Mudaliar, Justin Li, Daniel Lin et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text