EUVIMEDEUVIMED entdecken
Crossref

Atlas‐Augmented U‐Net Revisited: A Controlled Evaluation of Voxel‐Level Anatomical Priors for Data‐Efficient Brain Tumor Segmentation

Gaurav Dubey, Arti Sharma, Aatif Jamshed, Ruchi Jain, Abhishek Goyal

International Journal of Imaging Systems and Technology · 2026-11 · 36 (6)

DOI: 10.1002/ima.70474 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Imaging Systems and Technology, online erschienen am 09.11.2026. Verfasst von Gaurav Dubey und 4 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Anatomical atlases encode decades of neuroanatomical knowledge, and supplying that knowledge to a segmentation network is an appealing way to improve accuracy when labeled data are scarce. We ask whether a voxel‐level atlas prior delivers that improvement for 3D brain tumor segmentation, and we answer the question with a controlled ablation. An earlier version of this study added a Harvard‐Oxford parcellation to a segmentation network in two ways, region‐level tokenization and a voxel‐level fifth input channel, and reported that the voxel‐channel model (Atlas‐CNN, 5.6 M parameters) reached a mean Dice similarity coefficient (DSC) of 0.867 on a 200/51 training/test split of the Brain Tumor Segmentation 2023 (BraTS‐2023) dataset, roughly 3.4 points above a convolutional baseline. That comparison could not attribute the gain to anatomy: the baseline was a 0.3 M‐parameter network, about 18 times smaller than Atlas‐CNN, and no same‐backbone model with a non‐atlas fifth channel was trained. We therefore fixed the split and the 5.6 M U‐Net across every arm and varied only the fifth channel. On this single fixed split, the atlas channel reached a mean DSC of 0.9186, and no statistically significant difference was detected between the atlas and a parameter‐matched four‐channel baseline (0.9151, p = 0.896), a coordinate‐map channel (0.9159, p = 0.963), or a noise channel (0.9138, p = 0.426) by paired Wilcoxon signed‐rank tests, with negligible to small effect sizes ( r ≤ 0.11). A pre‐committed enhancing‐tumor gate, where anatomical context should matter most, likewise showed no significant atlas advantage over baseline ( p = 0.184) or coordinate ( p = 0.715). A random‐atlas control, a spatially coherent random partition, scored below every arm and significantly below the atlas (0.9046, p < 0.001, r > 0.46), which indicates that the network does respond to the fifth channel while deriving no measurable benefit from the atlas's anatomical content under these conditions. The 3.4‐point improvement reported earlier is therefore attributable to baseline construction. Because no separate validation set was held out, early stopping and checkpoint selection were performed on the test set for every arm; the absolute Dice values are therefore optimistic, although the paired between‐arm comparison remains controlled. These findings apply to the specific configuration evaluated here, one Harvard‐Oxford atlas supplied as a scalar channel, an affine registration pipeline, a U‐Net‐class backbone, and a single BraTS‐derived split without external validation, and should not be read as evidence that anatomical priors in general provide no benefit. We release the finding together with a control protocol (capacity, coordinate, and noise channels on a shared split) that any extra‐channel prior method should satisfy before an anatomical claim is made.

Zitieren

APA

Dubey, G., Sharma, A., Jamshed, A., Jain, R., & Goyal, A. (2026). Atlas‐Augmented U‐Net Revisited: A Controlled Evaluation of Voxel‐Level Anatomical Priors for Data‐Efficient Brain Tumor Segmentation. International Journal of Imaging Systems and Technology, 36(6). https://doi.org/10.1002/ima.70474

Vancouver

Dubey G, Sharma A, Jamshed A, Jain R, Goyal A. Atlas‐Augmented U‐Net Revisited: A Controlled Evaluation of Voxel‐Level Anatomical Priors for Data‐Efficient Brain Tumor Segmentation. International Journal of Imaging Systems and Technology. 2026;36(6). doi:10.1002/ima.70474

BibTeX
@article{Dubey2026,
  author = {Dubey, Gaurav and Sharma, Arti and Jamshed, Aatif and Jain, Ruchi and Goyal, Abhishek},
  title = {Atlas‐Augmented U‐Net Revisited: A Controlled Evaluation of Voxel‐Level Anatomical Priors for Data‐Efficient Brain Tumor Segmentation},
  journal = {International Journal of Imaging Systems and Technology},
  year = {2026},
  volume = {36},
  number = {6},
  doi = {10.1002/ima.70474}
}
Kostenlos starten →