Trained 3D U-Net for a 7-class microvessel segmentation

This 3D U-net was trained on self-annotated data of the public image database known as the "H01 release"

We trained a 3D semantic segmentation model, the 3D U-net, using our own 3D annotated images of capillaries extracted from the "H01 release" (https://h01-release.storage.googleapis.com/landing.html). This is a 7-class model with the following classes: (0) Background, (1) Basement Membrane, (2) Lumen, (3) Nuclei, (4) Mitochondria, (5) Endoplasmic Reticulum, and (6) Undetermined Structures. While the image data is available in 4 nanometer (nm) resolution in X-Y and 33 nm in Z, this model was trained on image volumes downsampled in X and Y by a factor of 2 (thus 8 nm resolution)