Class imbalance in diabetic retinopathy (DR) datasets, particularly the underrepresentation of advanced stages such as proliferative DR, limits classifier performance and necessitates specialized augmentation strategies. This study presents a data augmentation approach using StyleGAN2-ADAto generate stage-specific synthetic fundus images. Unlike conditional approaches requiring labels or segmentation masks, our method trains separate unconditional models for each DR severity stage, producing anatomically realistic images while preserving pathological characteristics. We propose a comprehensive evaluation framework comprising three components: quantitative assessment using distribution-based metrics (FID, KID, IS) and nearest-neighbor-based metrics (NN-SSIM, NN-PSNR); task-based evaluation through DR classification experiments with ConvNeXt Tiny, MobileNet V3 Small, and Xception architectures; and expert validation by three ophthalmologists assessing clinical plausibility. Classification experiments demonstrated improvements in overall accuracy, with ConvNeXt Tiny achieving 73.48% accuracy (+1.82% over baseline) when combining synthetic and classical augmentation. Inter-rater reliability analysis (κ = 0.723 between two evaluators) confirmed the clinical realism of generated images. Our unconditional StyleGAN2-ADA approach provides a scalable, mask-free augmentation solution addressing class imbalance in fundus imaging datasets, potentially enhancing DR screening systems in resource-constrained clinical settings.
Robot-centric elevation maps built from onboard sensing are always incomplete: occlusions, a limited field of view, and range limits leave large unobserved regions that traversability analysis and motion planning must still reason about. We present a supervised framework that completes these maps and reports a per-cell uncertainty. Its core is a ray-cone augmentation that removes angular sectors anchored at the sensor origin during training; unlike the random masks of image inpainting, these sectors match the coverage gaps of real deployments, such as camera failures or reduced camera configurations. Partial maps generated from four depth cameras along legged-robot trajectories in the TartanGround dataset are paired with dense ground truth, yielding 32,329 samples across five outdoor environments. An encoder–decoder network is trained with a masked -NLL loss and evaluated with a five-fold leave-one-environment-out protocol. The augmentation lowers the completion error on missing sensor sectors by 8.3 to 9.7%, depending on the sector width, at no measurable cost on uncorrupted partial inputs. The completed maps reach a hole root-mean-square error of 2.86 m, a 45% improvement over the strongest classical interpolation baseline.
Stacking ensembles are an effective means of improving prediction accuracy by combining several base models through a meta-model. However, in practice, the composition of base models is often formed intuitively, without a quantitative assessment of their mutual similarity and complementarity, which may lead to structural redundancy of the ensemble, increased computational costs, and reduced practical efficiency. This paper improves the method for selecting base models for a stacking ensemble using indicators of prediction and error diversity. The proposed approach involves evaluating model heterogeneity by simultaneously considering prediction correlation, error correlation, normalized error difference, and error covariance. This makes it possible to quantitatively assess model similarity and complementarity from several perspectives and to justify the formation of a more compact ensemble composition. The experimental validation was performed on a publicly available dataset for the task of predicting individual medical insurance costs. The results of the study showed that the reduced stacking model, built on the basis of three selected base models instead of the initial ten, provides an acceptable level of generalization ability close to that of the full model, while significantly reducing inference and training time. The obtained results indicate the practical feasibility of applying the proposed approach within the studied task to reduce the structural redundancy of stacking ensembles while preserving acceptable prediction quality and decreasing computational costs.