Ultrasound Confidence Maps with Neural Implicit Representation

Ultrasound Confidence Map (CM) is an image representation that indicates the reliability of the pixel intensity values presented within ultrasound B-mode images. Those maps are highly correlated with the probability of sound reaching specific depths. Commonly, CM is calculated based only on the B-mode images, without anatomy awareness. Without clear anatomical landmarks or contextual information, CMs might misrepresent the certainty of features detected within the ultrasound images. We propose a novel deep-learning approach for CM calculation that is specific to the anatomy and based on physical principles of echo propagation. We rely on the physics-inspired intermediate representation maps of Ultra-Nerf to compute CMs with observation-angle awareness, similar to the clinical practice. Our method outperforms other methods on downstream tasks such as shadow segmentation and compounding. Additionally, we open-source the code and a tracked ultrasound dataset to promote more research in this direction at https://github.com/MrGranddy/Redefining-Confidence-Maps.