AI-based Automatic segmentation of 3D ultrasound images of the prostate
The objective is to produce automatically contours of the prostate from 3D ultrasound images. The 3D volume is represented as a stack of 2D slices. The approach is based on a combination of convolutional networks: 3 networks specialize in the segmentation of 2D slices in a given direction (axial, sagittal, frontal) while the 4^th network develops an ensemble learning strategy; for each voxel of the volume, it determines the optimal weighted combination of the three segmentations by the direction-specific networks. The rationale is that some regions of the organ are more visible on a specific viewing direction (for instance the apex of the prostate is more visible on sagittal slices). The method has been trained and evaluated on a clinical database from the Grenoble University Hospital (in agreement with French regulation on clinical data).
Reference: Clément Beitone and Jocelyne Troccaz. Multi-expert fusion: An ensemble learning framework to segment 3D TRUS prostate images. Medical Physics, 2022:49(8):5138-48. https://hal.archives-ouvertes.fr/hal-03654488 .
Funding: Region “Rhône-Alpes Auvergne” project PRONAVIA, PSPC DIANA from BPI France and MIAI AI Institute
Collaborations: Clinical partners and KOELIS company.
Status: IN PROGRESS Work with Clément BEITONE, TIMC
Contact : jocelyne.troccaz@univ-grenoble-alpes.fr and clement.beitone@univ-grenoble-alpes.fr