@

초록

Deep models for retinal fundus analysis often depend on handcrafted preprocessing (e.g., CLAHE, background division) that is dataset-specific and costly to tune. We propose a Frequency Modulating Network (FMN), a parameter-efficient, plug-and-play module that operates in the Fourier domain using soft radial band masks (Gaussian/RBF) and a residual, mean-preserving gate. FMN can be placed either (i) before the backbone as a learnable, end-to-end preprocessing block (image space) or (ii) after high-level stages as a feature-refinement block (feature space). On the APTOS 2019 five-grade diabetic retinopathy benchmark, across CNN (ResNet-18/50) and ViT backbones, FMN consistently improves accuracy, specificity, sensitivity, F1, and macro-AUC while adding only a single FFT/iFFT pair per insertion and negligible parameters. An ablation on the number of radial bands shows a U-shaped trend, with moderate band counts performing best. A comparison with hand-crafted pipelines (stochastic augmentations, CLAHE, Ben-Graham) finds FMN competitive or superior without task-specific tuning. Notably, an input-conditioned variant - where band gains are generated from the input image - underperforms the unconditioned FMN, suggesting that per-image modulation can overfit sample-specific quirks, whereas the unconditioned FMN learns a global task/domain-level spectral regularizer that generalizes more robustly.

키워드