An artificial-intelligence model for coronary artery segmentation in X-ray angiography reported the highest Dice and IoU results on the MOSXAV test set and the highest Dice, IoU and recall in an external XACV evaluation. It did not lead every measure: AdapterSeg had the highest recall on MOSXAV test data, while nnWNet had slightly higher precision on XACV.
The document is an arXiv version 1 preprint posted on 25 August 2026. Its central comparison covers the proposed method and other segmentation systems on MOSXAV, followed by an evaluation on XACV without fine-tuning.
The results depended on the split
On MOSXAV validation, which was used for model selection, the method reported Dice 84.75, IoU 74.05, precision 85.22 and recall 85.97. Those were the best reported Dice, IoU and recall values among the listed methods. Precision was the exception: nnU-Net and nnWNet recorded higher values.
The MOSXAV test set was used for evaluation. There, the reported scores were Dice 53.47, IoU 39.09, precision 46.29 and recall 77.74. The proposed method had the best Dice and IoU and the highest precision in that comparison, but AdapterSeg had the highest recall.
Taken together, the two MOSXAV comparisons did not produce one uniform winner across all four measures. The proposed method led Dice and IoU on both splits, led recall on validation, and led precision on the test set, while other methods topped the remaining measures.
The cross-dataset check
All models were trained on MOSXAV and evaluated on XACV without fine-tuning. That design provided an external comparison using the same trained systems on a different dataset.
On XACV, the proposed method reported Dice 70.25, IoU 54.95, precision 76.84 and recall 65.37. It had the best Dice, IoU and recall, while nnWNet had slightly higher precision.
Across the three reported evaluations, the method’s ranking was most consistent for Dice and IoU: it led both measures on MOSXAV validation, MOSXAV test and XACV. Its recall lead appeared on validation and XACV, while its precision lead appeared only on the MOSXAV test.
What changed in the model
The proposed system is a mixture-of-experts feature adapter. In the paper’s architecture, one adapter branch is replaced with multiple lightweight experts and a router.
In the reported implementation, there were four experts, with two activated for each feature. MoE adapters were inserted into all ViT stages, forming the sparse top-k setup examined in the study.
The models were trained with AdamW for 40 epochs, and selection was based on validation Dice. MOSXAV training data were used for training, validation data for model selection, and test data for evaluation.
A routing choice at the center
In the MOSXAV ablation, the study compared routing strategies. The top-k variant reported Dice 84.75, IoU 74.05, precision 85.22 and recall 85.97, with 356M parameters.
Among those routing strategies, top-k routing had the best reported Dice, IoU, precision and recall. The result supports the top-k configuration as the strongest-performing routing variant within this ablation.
How much certainty the figures carry
The reported figures are point scores, and no inferential uncertainty measures are given for the validation, test or external comparisons. The analysis presents these as descriptive comparisons of model metrics.
The work was supported by EPSRC UK under grant EP/X023826/1. The authors declare no relevant competing interests.
Paper data and sources
Original title: MoE-based Feature Adapter for Prompt-free Binary Coronary Artery Segmentation in X-ray Angiography
Authors: Lin Xi, Yingliang Ma
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text