Segmentation From Partial Views Via Patient-Specific Shape Priors
Precision surgical interventions rely on accurate integration of preoperative and intraoperative imaging to guide clinical decision-making and improve patient outcomes. Traditionally, most 3D imaging modalities capture the entirety of the target object, allowing for the segmentation of its entire volume. However, some medical imaging devices trade full field-of-view for other considerations, such as size and ability to access the target organ in unique ways. This is particularly true of ultrasound, where imaging deeper organs from outside the body is not viable due to attenuation, thus motivating the need for alternative imaging strategies. In scenarios such as image-guided intervention, it is often necessary to put this partial view of the object in its full anatomical context, such as alignment to a pre-operative MRI. We propose a method for jointly segmenting the portion of the object visible in the acquired image and estimating its full shape. We do this by combining a pre-trained shape prior with a patient-specific expectation of organ shape acquired via pre-operative imaging. In a simulated dataset for prostate MRI/US fusion, we show the ability to accurately estimate the prostate shape from ultrasound images capturing only a fraction of its total volume.