EUVIMEDEUVIMED entdecken
Crossref

Morphological Transition Flow-based Feature Extraction for Five-Class Diabetic Retinopathy Classification

Basma Esserkassi, Souad Eddarouich, Abdennaser Bourouhou

Journal of Innovative Image Processing · 01.07.2026 · 8 (3) · 810-826

DOI: 10.36548/jiip.2026.3.004 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Innovative Image Processing, online erschienen am 01.07.2026. Verfasst von Basma Esserkassi und 2 weiteren Autor:innen. Die Arbeit nennt 24 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Diabetic retinopathy threatens the vision of 103 million people globally, yet computational and infrastructural barriers continue to block automated screening where it is needed most. We propose RF-MTF, a three-module framework evaluated on 34,860 EyePACS fundus images across five ICDR grades (validation set n=6,972). Module 1 applies AMS-CLAHE with Gabor morphological guidance. Module 2 introduces the Morphological Transition Flow — RGB channels processed by random projection networks (ELM, RVFL, BLS) with dual-scale morphological operations, yielding 384-dimensional feature vectors. Module 3 benchmarks three classifiers across 34 configurations spanning four optimization levels. Preprocessing variants produced near-identical F1-scores (0.9045–0.9086, Δ=0.41 pp) despite large image quality divergence, questioning the assumption that preprocessing optimization is critical for DR classification. The optimal configuration reached F1-weighted=0.9701, F1-macro=0.8910, AUC-ROC=0.9941. Random Forest achieved exceptional configuration-wise stability (σ=0.0002), in contrast to XGBoost's high optimization sensitivity (σ=0.2097) and SVM's persistent low performance on this feature space (σ=0.0108, F1≈0.25 across all configurations). Inference runs in 1–3 ms within a 20–40 MB footprint. RF-MTF delivers competitive 5-class ICDR classification on CPU hardware (F1-weighted=0.9701, F1-macro=0.8910), with minority-class performance (Grade 4 F1=0.78, 95% CI [0.70, 0.85], n=70) reflecting the inherent statistical challenges of evaluating rare classes under the natural 95.9:1 class imbalance of real-world DR screening populations — identified as the primary minority-class improvement target. External validation on Messidor-2, APTOS 2019, and IDRiD is the immediate next step.

Zitieren

APA

Esserkassi, B., Eddarouich, S., & Bourouhou, A. (2026). Morphological Transition Flow-based Feature Extraction for Five-Class Diabetic Retinopathy Classification. Journal of Innovative Image Processing, 8(3), 810-826. https://doi.org/10.36548/jiip.2026.3.004

Vancouver

Esserkassi B, Eddarouich S, Bourouhou A. Morphological Transition Flow-based Feature Extraction for Five-Class Diabetic Retinopathy Classification. Journal of Innovative Image Processing. 2026;8(3):810-826. doi:10.36548/jiip.2026.3.004

BibTeX
@article{Esserkassi2026,
  author = {Esserkassi, Basma and Eddarouich, Souad and Bourouhou, Abdennaser},
  title = {Morphological Transition Flow-based Feature Extraction for Five-Class Diabetic Retinopathy Classification},
  journal = {Journal of Innovative Image Processing},
  year = {2026},
  volume = {8},
  number = {3},
  pages = {810--826},
  doi = {10.36548/jiip.2026.3.004}
}
Kostenlos starten →