Table 1.
Demographic data.
Fig 1.
Study overview to improve 2D MRI visualization of rotator cuff tendons using segmentation network.
In this study, training segmentation was conducted twice in order to improve the segmentation result of the cuff tendons, with A) secondary labelling process and B) refined segmentation training. In the secondary labelling process, a secondary labelled dataset was constructed based on the false-positive segmentation results for the cuff tendon of the segmentation model using the manually annotated dataset (4 classes; muscle, humerus, cuff tendon, cartilage), and the newly formed 4 + 1 class dataset was trained once more to obtain a refined segmentation result.
Fig 2.
Comparison between 2D and 3D convolutional calculation.
The number of dimensions represents the number of directions in the filter.
Fig 3.
Pipeline of the nnU-Net segmentation model of 2D rotator cuff MRI.
Fig 4.
Process of adapting secondary labelling.
To generate a complete secondarily-labelled data set (4+1 class), we split the data set into 5 folds and used 5 segmentation models trained by each fold.
Fig 5.
Training and validation loss curve of 3D semantic segmentation of the rotator cuff.
Table 2.
Segmentation result for each label in 2D and 3D U-Net according to the secondary labelling effect.
Fig 6.
Reduction of false-positive segmentation results using secondary labelling.
(a) The result of prediction using 3D U-Net. (b) The result of adapting secondary labelling. (c) Ground truth.
Fig 7.
Segmentation performance according to cuff tear size.