Strojové učení pro automatickou segmentaci retinálních hemoragií

Abstract

Diabetic retinopathy is among the leading causes of vision loss in the working-age population. Early detection of retinal hemorrhages in fundus photograph screening can significantly improve patient prognosis; however, manual image assessment is time-consuming and subject to subjectivity. This work focuses on the development and validation of automated segmentation models for retinal hemorrhage detection using deep learning. Three neural network architectures (U-Net/ResNet34, DeepLabV3Plus/ResNet50, and Attention U-Net/ResNet34) were compared on a dataset of 130 fundus images annotated by three independent ophthalmologists. The STAPLE algorithm was used to eliminate inter-rater variability and create consensus ground truth. Twelve experiments were conducted evaluating the influence of annotation quality, model architecture, and training parameters on segmentation performance under extreme class imbalance. Results demonstrate the importance of ground truth quality, with models trained on STAPLE consensus achieving Dice coefficients up to 0.744, representing a 12% improvement over models trained on single-physician annotations. Cross-doctor validation revealed a 5–10% performance decrease when testing on a different annotator, confirming the risk of overfitting to specific annotation styles. The best model achieved very strong correlation between predicted and actual hemorrhage extent with low mean absolute error. This work provides a methodology for developing clinically applicable segmentation models and demonstrates that consensus ground truth is critical for achieving robust results. The developed models can serve as support tools for diabetic retinopathy screening.