EUVIMEDEUVIMED entdecken
Crossref

Information technology for biomedical image processing based on transfer learning for ophthalmological pathology classification

A.Yu. Shegera, O.M. Velychko, О.М. Datsok

Radiotekhnika · 25.07.2026 · (225) · 115-121

DOI: 10.30837/rt.2026.2.225.11 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Radiotekhnika, online erschienen am 25.07.2026. Verfasst von A.Yu. Shegera und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The paper solves an urgent scientific and practical problem of improving the efficiency of automated diagnostics of ophthalmological diseases by developing a specialized information technology for processing biomedical visual signals. The aim of the study is to create a convolutional neural network architecture capable of reliable multi-label classification of pathologies using fundus eye images under conditions of limited and highly imbalanced datasets. A comprehensive analysis of the ODIR-5K database was performed, which revealed an extreme class imbalance and the presence of technical artifacts. To overcome these problems, preliminary data cleaning and an algorithm for mapping text diagnoses into standardized target class vectors were proposed. The ResNet50 architecture with a transfer learning strategy was used as a basis, which made it possible to compensate for the excess model capacity and prevent severe overfitting inherent in classical networks when working with medical data. A modified hybrid Focal Loss function was developed and integrated, which contains additional clinically justified penalties for an excessive number of predicted classes and for illogical combinations of diagnoses (for example, simultaneous prediction of normal and pathological states). To optimize the training process, the AdamW algorithm with decoupled weight decay and a two-stage parameter tuning strategy (Feature Extraction and Fine-Tuning) with dynamic learning rate changes were used. Real-time data augmentation methods were applied. Performance evaluation on the test sample confirmed a significant increase in quality: the average Macro Avg F1-score increased more than threefold compared to the baseline model, and the area under the PR-curve for key pathologies increased several times. Visualization of the decision-making process using the Grad-CAM algorithm proved that the developed transfer architecture is able to focus on clinically relevant anatomical structures, such as exudates or hemorrhages, successfully ignoring technical shooting artifacts. The obtained results indicate a potential of the proposed information technology for use as an auxiliary decision support tool or a preliminary screening system to optimize the work of an ophthalmologist. Further research will be aimed at integrating visual transformers and generative adversarial networks to compensate for the lack of data on rare diseases.

Zitieren

APA

Shegera, A., Velychko, O., & Datsok, О. (2026). Information technology for biomedical image processing based on transfer learning for ophthalmological pathology classification. Radiotekhnika, 115-121. https://doi.org/10.30837/rt.2026.2.225.11

Vancouver

Shegera A, Velychko O, Datsok О. Information technology for biomedical image processing based on transfer learning for ophthalmological pathology classification. Radiotekhnika. 2026:115-121. doi:10.30837/rt.2026.2.225.11

BibTeX
@article{Shegera2026,
  author = {Shegera, A.Yu. and Velychko, O.M. and Datsok, О.М.},
  title = {Information technology for biomedical image processing based on transfer learning for ophthalmological pathology classification},
  journal = {Radiotekhnika},
  year = {2026},
  number = {225},
  pages = {115--121},
  doi = {10.30837/rt.2026.2.225.11}
}
Kostenlos starten →