Automatic segmentation of brachytherapy seeds in 3D ultrasound images
In order to allow dynamic and adaptive, robotic or manual brachytherapy, it is essential to be able to robustly detect the real position of the deposited grains. In addition to the intrinsic difficulties linked to segmentation in ultrasound imaging, there are the specificities linked to the grains (smallness of the grains and very many artefacts they produce - see opposite above). We have developed an approach using a prior detection of needles via a Bayesian classification method to infer a search area. The grains are detected thanks to the registration of a geometric model of the grain train with the image segmented by Bayesian classification in the search area. Then for each grain, the position is then refined by iconic registration of a grain appearance model in the images with the sub-volume concerned. This approach has been evaluated on phantoms (see opposite, bottom) and on patient data.
References : Hatem Younes, Jocelyne Troccaz, and Sandrine Voros. Machine learning and registration for automatic seed localization in 3D US images for prostate brachytherapy. Medical Physics, 48(3):1144-1156, January 2021. https://hal.archives-ouvertes.fr/hal-03023560 . Younes, H., Voros, S., & Troccaz, J. (2018, April). Automatic needle localization in 3D ultrasound images for brachytherapy. In 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018) (pp. 1203-1207). IEEE. https://hal.archives-ouvertes.fr/hal-01800108/document PhD thesis of Hatem Younes, 2020, in French (https://hal-emse.ccsd.cnrs.fr/INPG/tel-03009135v1 )
Funding: ANR FOCUS project (coordinated by LATIM, Brest, France)
Collaborations : Partners of the FOCUS project (LATIM lab, Brest University Hospital, Grenoble University Hospital, Koelis company)
Status : FINISHED
Contact : jocelyne.troccaz@univ-grenoble-alpes.fr