METHOD OF SEGMENTING NEGATIVE VOLUMES IN COMPLEX 3D STRUCTURES

The present disclosure is related to the field of 3D image segmentation, in particular to a method of automatically determining the contours of the constituent parts of complex (compound) structures on 3D images using technique of segmenting negative volumes and deep neural network. The technical result of the claimed invention consists in providing faster and more accurate negative volume segmentation of the compound 3D structures of living and non-living objects. The claimed result is achieved due to implementation of the method, including the following steps: obtaining 3D image data of the compound 3D structure; selecting a volume of interest, comprising a negative volume on the obtained 3D image data of the compound 3D structure, wherein the negative volume is an inner space between at least two adjacent components of the compound 3D structure; segmenting the adjacent components of the compound 3D structure; 3D reconstructing the segmented adjacent components; segmenting the negative volume, wherein segmentation of the negative volume including at least the following: mesh inflating an inner surface of the first reconstructed component to fit into an inner surface of the second reconstructed component until the entire negative volume is occupied, and obtaining the inflated first reconstructed component, wherein surface mesh inflation including at least the following: obtaining vertices of the first and the second reconstructed components; obtaining number of vertices of the first and the second reconstructed components; obtaining faces of the first and the second reconstructed components based on obtained vertices and number of vertices; obtaining normal vector to each face of the first reconstructed component; obtaining coordinates of each vertex of the first reconstructed component; while the first reconstructed component does not intersect with the second reconstructed component extruding each face of the first reconstructed component separately along local normal and shifting coordinates of each vertex of the first reconstructed component increasing parameters, comprising length, radius, depth; subtracting the first reconstructed component from the inflated first reconstructed component; 3D reconstructing the inner space between the adjacent components of the compound 3D structure. La présente divulgation, qui relève du domaine de la segmentation d'images 3D, concerne en particulier un procédé de détermination automatique des contours des parties c