A beamformer-independent method to predict photoacoustic visual servoing system failure from a single image frame

Visual servoing is a robotic method that has the potential to assist surgeons with tracking tool tips when attached to optical fibers to create photoacoustic images that are autonomously monitored. Currently, this approach must be tested with multiple image frames and multiple laser energies prior to each surgery in order to identify the minimum required energy that will not cause system failure over the number of frames tested. This study investigates possible integration of the generalized contrast-to-noise ratio (gCNR) into pre-surgical procedures as a method to predict system failure from only a single image frame. Photoacoustic data were acquired from an optical fiber inserted in a plastisol phantom or in the left ventricle of an in vivo swine heart. Raw data were processed with delay-and-sum (DAS) and short-lag spatial coherence (SLSC) beamforming (M = 25). gCNR values were estimated from a 3 mm x 3 mm region of interest (ROI) surrounding the photoacoustic target coordinates provided by the visual servoing algorithm. The prediction function modelled from phantom data was fit with R2 values of 0.992 and 0.991 for DAS and SLSC beamformers, respectively. When applying this fit to the in vivo test data, the RMSE between measured segmentation accuracy and the prediction functions was 9.34% for DAS images and 4.78% for SLSC images. These results indicate that the newly introduced image quality metric gCNR has sufficient robustness to predict the performance of visual servoing segmentation tasks and thereby mitigate the burden, time, and requirements of testing multiple image frames prior to the initiation of a surgery.