Abstract: Existing uncertainty-guided learning frameworks typically rely on passive pixel selection strategies and struggle to effectively distinguish aleatoric uncertainty arising from physical noise from the model’s epistemic uncertainty, resulting in excessive allocation of sampling resources to unlearnable noise regions. This paper proposes a multi-dimensional calibrated active probing (MC-AP) framework for instance-level segmentation of anatomical structures in ultrasound images. The framework introduces an attention-based sampling head with a learnable query mechanism, transforming traditional passive, non-differentiable pixel selection into an end-to-end differentiable active probing process in the feature space, thereby enabling the co-evolution of the sampling strategy and the segmentation network. To filter out noise interference within mixed uncertainties, the active probing mechanism is calibrated across multiple dimensions: in the physical dimension, an image-gradient-guided structural prior constraint is introduced to force the probes to focus on true anatomical boundaries; in the probabilistic and performance dimensions, the confidence of the predictive distribution is explicitly modeled, and a difficulty-enhanced sampling reward function is designed to keep the optimization process focused on cognitively difficult regions with large persistent prediction errors. In addition, a spatial diversity constraint is imposed to ensure complete coverage of complex anatomical structures by the sampling points. Experimental results on challenging ultrasound datasets, including MEIS and TN3K, show that MC-AP achieves mask mean average precision (Mask-mAP) values of 57.01% and 51.17%, respectively. These results indicate that combining active sampling with multi-dimensional calibration helps mitigate noise interference, locate difficult regions, and improve ultrasound image instance segmentation performance.