Evidential Uncertainty-Guided Interactive Segmentation of Ultrasound Images
An end-to-end interactive medical-image segmentation paradigm that uses evidential deep learning (Dempster-Shafer theory and Subjective Logic) to estimate pixel-wise predictive uncertainty, then automatically directs interaction prompts to the highest-uncertainty regions to reduce the number of prompts and iterations a user must supply. A trainable calibration mechanism refines the certain/uncertain boundary to make the uncertainty estimates more trustworthy, yielding efficient, specialized segmentation of noisy ultrasound images.
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. XX, NO. XX, XXXX 2020 1 Evidential Calibrated
EUGIS (Evidential Uncertainty-Guided Interactive Segmentation) is an interactive segmentation method for medical ultrasound images. Instead of relying on many manually placed or randomly sampled poin…