Title: Implicit Fuzzification via Bounded Noise Injection for Robust Medical Image Segmentation

Abstract:

Image segmentation is fundamentally constrained by boundary ambiguity, which stems from sampling-induced information loss and the inherent uncertainty associated with pixel-wise labeling. While encoder-decoder models like U-Net deliver strong performance, they frequently generate overconfident predictions that overlook the ambiguity present in transition regions. To mitigate this challenge, we introduce \textbf{NoiseUNet}, a straightforward yet effective framework designed to inject bounded perturbations into skip connections, thereby regularizing cross-scale feature fusion. This approach enhances robustness against local feature fluctuations and encourages the development of boundary-aware representations. Theoretically, these perturbations create an implicit fuzzification effect, producing soft, data-driven memberships without the need for explicit fuzzy modeling. Additionally, we present \textbf{ThyR}, a real-world thyroid ultrasound dataset characterized by inherently ambiguous boundaries. Our experimental results show that NoiseUNet consistently boosts both segmentation accuracy and boundary fidelity.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC