EUVIMEDEUVIMED entdecken
Crossref

DFS ‐Net: Dual‐Frequency Perception and Spatially Adaptive Loss for Brain Tumor Segmentation

Jingjun Dong, Cuifang Gao

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

DOI: 10.1002/ima.70475 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Imaging Systems and Technology, online erschienen am 10.11.2026. Verfasst von Jingjun Dong und 1 weiteren Autor:innen. Die Arbeit nennt 61 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate brain tumor segmentation requires boundary‐sensitive local detail modeling and stable global semantic representation. Existing convolutional, transformer‐based, and foundation‐model approaches have improved segmentation accuracy, but many of them still process frequency‐dependent cues implicitly, making it difficult to balance high‐frequency tumor boundaries with low‐frequency anatomical context. In addition, commonly used overlap losses do not explicitly emphasize boundary‐adjacent pixels or hard samples under severe class imbalance. To address these limitations, we propose DFS‐Net, an end‐to‐end dual‐frequency segmentation framework composed of a Hybrid Frequency Perception Unit (HFPU), Frequency‐Spatial Attention (FSA), and Adaptive Residual Block (ARB). HFPU separates detail‐ and context‐oriented feature streams, FSA lets high‐frequency features query low‐frequency semantic cues, and ARB adaptively fuses attention‐enhanced and residual information. We further introduce Spatial‐Adaptive Unified Focal Loss (SA‐UFL), which combines spatial boundary weighting with adaptive difficulty modulation. On the LGG and BraTS 2021 datasets, DFS‐Net achieves Dice scores of 0.9310 and 0.94/0.87/0.85 for WT/TC/ET, corresponding to a 1.3 percentage‐point gain over EfficientNetB4 on LGG, matching the strongest reported WT Dice on BraTS 2021, and improving TC/ET by 1.0/1.0 percentage points over the strongest reported BraTS baselines. Additional validation on BraTS 2023 Adult Glioma and BraTS‐SSA 2025 yields macro Dice scores of 0.9010 and 0.7964, respectively, the highest among the compared 2D models.

Zitieren

APA

Dong, J., & Gao, C. (2026). DFS ‐Net: Dual‐Frequency Perception and Spatially Adaptive Loss for Brain Tumor Segmentation. International Journal of Imaging Systems and Technology, 36(6). https://doi.org/10.1002/ima.70475

Vancouver

Dong J, Gao C. DFS ‐Net: Dual‐Frequency Perception and Spatially Adaptive Loss for Brain Tumor Segmentation. International Journal of Imaging Systems and Technology. 2026;36(6). doi:10.1002/ima.70475

BibTeX
@article{Dong2026,
  author = {Dong, Jingjun and Gao, Cuifang},
  title = {DFS ‐Net: Dual‐Frequency Perception and Spatially Adaptive Loss for Brain Tumor Segmentation},
  journal = {International Journal of Imaging Systems and Technology},
  year = {2026},
  volume = {36},
  number = {6},
  doi = {10.1002/ima.70475}
}
Kostenlos starten →