EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

CRA-Net: A Wavelet-Decomposed Vision Transformer With Cross-Radiomics Attention for Fine-Grained Biomedical Image Classification

Sifeddine Elkardoudi, Ahmed Eddaoui, Mohamed Talea

IEEE Access · 2026 · 14 · 125864-125878

DOI: 10.1109/access.2026.3723838 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Sifeddine Elkardoudi und 2 weiteren Autor:innen. Die Arbeit nennt 41 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Fine-grained biomedical image classification requires networks that can separate cell phenotypes which look nearly identical but carry opposite pharmacological meaning. In high-throughput drug screening (HTS), a false positive is far costlier than a false negative: it pushes an inert compound into weeks of downstream validation. Standard CNNs and Vision Transformers score well on accuracy and AUC yet still produce false positives, because they learn texture from pixels alone and miss the quantitative shape cues a cell biologist would use. We introduce CRA-Net (Cross-Radiomics Attention Network), a modular framework that couples wavelet-domain visual encoding with 102 hand-crafted PyRadiomics descriptors through a dedicated cross-attention fusion block. CRA-Net ingests dual-channel fluorescence microscopy (GFP and DAPI), gates the peri-nuclear cytoplasm via a DAPI-driven spatial mask, splits the gated signal into low- and high-frequency sub-bands with a differentiable Haar wavelet transform, and encodes both streams through interleaved Vision Transformer blocks with cross-frequency attention. On the primary stratified 80/20 evaluation (234 bags derived from 1,580 inflammasome speck images, 9 independent compound groups), CRA-Net reached 95.74% bag-level accuracy, $0.9500~\text {F1}$ , $0.9725~\text {ROC AUC}$ , and a precision of 1.0000 using only 5.86M trainable parameters, representing a $23\times $ reduction over VGG-16. CRA-Net produced zero false positives on these 47 validation bags; we emphasise that this is an observation on a small cohort and not a guarantee of perfect precision, since the Wilson 95% confidence interval for the precision point estimate spans $[{0.83, 1.00}]$ and overlaps several baselines. A strict 5-fold Stratified Group Cross-Validation splitting by compound identity confirms that no compound information crosses fold boundaries; it also shows markedly lower and highly variable performance (mean accuracy $0.612\pm 0.377$ , mean $\text {F1} 0.607\pm 0.418$ ), so the 80/20 result should not be read as evidence of generalisation to unseen compounds. Our conclusions are scoped to this single inflammasome speck assay, for which structured radiomics fusion delivers a measurable reduction in false discoveries; extension to other phenotypic endpoints and to broader biomedical screening remains to be established.

Zitieren

APA

Elkardoudi, S., Eddaoui, A., & Talea, M. (2026). CRA-Net: A Wavelet-Decomposed Vision Transformer With Cross-Radiomics Attention for Fine-Grained Biomedical Image Classification. IEEE Access, 14, 125864-125878. https://doi.org/10.1109/access.2026.3723838

Vancouver

Elkardoudi S, Eddaoui A, Talea M. CRA-Net: A Wavelet-Decomposed Vision Transformer With Cross-Radiomics Attention for Fine-Grained Biomedical Image Classification. IEEE Access. 2026;14:125864-125878. doi:10.1109/access.2026.3723838

BibTeX
@article{Elkardoudi2026,
  author = {Elkardoudi, Sifeddine and Eddaoui, Ahmed and Talea, Mohamed},
  title = {CRA-Net: A Wavelet-Decomposed Vision Transformer With Cross-Radiomics Attention for Fine-Grained Biomedical Image Classification},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {125864--125878},
  doi = {10.1109/access.2026.3723838}
}
Kostenlos starten →