Nikita Vladimirov1 and Ekaterina A. Brui1
1Department of Physics and Engineering, ITMO University, Saint Petersburg, Russian Federation
1Department of Physics and Engineering, ITMO University, Saint Petersburg, Russian Federation
U-Net convolutional neural networks were used for automatic wrist cartilage segmentation. Utilisation of a truncated U-Net archutecture and data augmentation allowed to increase the segmentation accuracy especially in lateral slices in comparison to previously published results.
Illustrations of performance (with medium accuracy) of the U-Net and truncated U-Net on several slices from one 3D image (green: correctly segmented pixels [true positives]; blue: pixels incorrectly assigned to the background [false negatives]; and red: pixels incorrectly assigned to the cartilage [false positives]).
Results for layer-to-layer analysis: average DSC values for the zones with different percentage of cartilage within the 3D images; average 3D DSC; average DSC for medial slices; the time needed for the segmentation of one slice. *The network was tested on a PC with common characteristics (an Intel Core i5-7640X processor with 32 Gb of RAM).