PI-Controlled Uncertainty for Steady-State Error Elimination in Ultrasound Image Segmentation

ICLR 2026 Conference Submission14922 Authors

Keywords: Medical Image Segmentation, Uncertainty, Control Theory, PI Controller

Abstract: Accurate segmentation of anatomical structures from medical ultrasound images is essential for reliable diagnosis, yet conventional training losses often leave persistent steady-state errors, especially along ambiguous boundaries. These losses act as control variables generated by a proportional controller, since they respond only to instantaneous discrepancies and lack the memory required to correct long-term deviations. To overcome this limitation, we rethink segmentation training as a closed-loop control system where uncertainty acts as the control variable. Building on this perspective, we introduce a proportional–integral (PI) control mechanism that integrates both present and historical error signals into the optimization process, enabling the model to systematically eliminate steady-state errors and deliver sharper, more reliable boundary predictions. Unlike existing uncertainty-based approaches that rely solely on fixed loss terms, our method provides a principled mechanism to incorporate dynamic feedback into training. The framework is model-agnostic and introduces no additional inference overhead, making it directly compatible with real-time segmentation backbones. Extensive experiments on clinical medical ultrasound datasets demonstrate consistent improvements over state-of-the-art baselines. These results confirm that our framework offers an effective solution for eliminating steady-state errors in medical ultrasound image segmentation under challenging conditions. Our code is available at https://anonymous.4open.science/r/PI-control-uncertainty-B82C.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 14922