EUVIMEDEUVIMED entdecken
Studienprotokoll

HDFF-Net: A Hybrid Dual-Feature Fusion Network with Bidirectional Gated Fusion for Colposcopic Transformation Zone Classification, Evaluated under a Leakage-Audited Cluster-Level Protocol

B Shubhaker, B S Raghavendra

Journal of Imaging Informatics in Medicine · 09.11.2026

DOI: 10.1007/s10278-026-02497-8 ↗

Kurz gefasst

Studienprotokoll in der Zeitschrift Journal of Imaging Informatics in Medicine, online erschienen am 09.11.2026. Verfasst von B Shubhaker und 1 weiteren Autor:innen. Beteiligte Einrichtung: Department of Electronics and Communication Engineering, National Institute of Technology Karnataka, Surathkal, Karnataka, 575025, India. shubhaker.s…. Die Arbeit nennt 76 Literaturangaben.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Cervical cancer screening through colposcopy depends critically on accurate classification of the transformation zone (TZ) according to the 2011 IFCPC nomenclature, as TZ type directly governs treatment eligibility (ablative versus excisional). Inter-rater agreement among trained colposcopists for TZ-type assignment is only κ ≈ 0.42-0.60, well below the κ ≥ 0.80 threshold considered reliable for clinical decision-making. Existing automated methods rely either on handcrafted feature pipelines or deep transfer learning in isolation, and published performance figures on public colposcopy atlases have not been systematically audited for evaluation integrity. This paper reports a leakage-audited, cluster-level evaluation of a hybrid handcrafted-plus-deep fusion design on a public colposcopy atlas and asks whether the fusion adds value over simpler baselines when only a few hundred images are available. We evaluate HDFF-Net (Hybrid Dual-Feature Fusion Network), a dual-stream design that combines a 10,322-dimensional multi-scale handcrafted feature branch (MS-GLCM, MS-LBP, MS-HOG), projected to 512 dimensions and recalibrated by a 1D Squeeze-and-Excitation (SE) module, with an EfficientNet branch augmented by CBAM. The two streams are joined by a bidirectional gated fusion (BGF) module that uses feature-wise sigmoid gates; it is not an attention mechanism (no query-key interaction, softmax, attention heads or spatial tokens). Models are trained on a stratified 70/15/15 split made at the level of 194 unique image clusters (seed = 42), obtained by perceptual-hash deduplication (pHash, Hamming distance

Schlagwörter

Stichwörter der Autor:innen

Cervical cancer screeningColposcopyData leakageDeep learningEfficientNetB0Gated feature fusionIFCPC nomenclatureMcNemar testMulti-scale HOGPerceptual-hash deduplicationSmall-sample evaluationSqueeze-and-excitationTransformation zone classification

Autor:innen und Institutionen

  • Department of Electronics and Communication Engineering, National Institute of Technology Karnataka, Surathkal, Karnataka, 575025, India. shubhaker.subbu@gmail.com.
  • Department of Electronics and Communication Engineering, National Institute of Technology Karnataka, Surathkal, Karnataka, 575025, India.

Zitieren

APA

Shubhaker, B., & Raghavendra, B. S. (2026). HDFF-Net: A Hybrid Dual-Feature Fusion Network with Bidirectional Gated Fusion for Colposcopic Transformation Zone Classification, Evaluated under a Leakage-Audited Cluster-Level Protocol. Journal of Imaging Informatics in Medicine. https://doi.org/10.1007/s10278-026-02497-8

Vancouver

Shubhaker B, Raghavendra BS. HDFF-Net: A Hybrid Dual-Feature Fusion Network with Bidirectional Gated Fusion for Colposcopic Transformation Zone Classification, Evaluated under a Leakage-Audited Cluster-Level Protocol. Journal of Imaging Informatics in Medicine. 2026. doi:10.1007/s10278-026-02497-8

BibTeX
@article{Shubhaker2026,
  author = {Shubhaker, B and Raghavendra, B S},
  title = {HDFF-Net: A Hybrid Dual-Feature Fusion Network with Bidirectional Gated Fusion for Colposcopic Transformation Zone Classification, Evaluated under a Leakage-Audited Cluster-Level Protocol},
  journal = {Journal of Imaging Informatics in Medicine},
  year = {2026},
  doi = {10.1007/s10278-026-02497-8}
}
Kostenlos starten →