Fig 1.
The schematic diagram of U-Net.
(A) U-Net encoding-decoding process. In the encoding stage the dimensions are reduced. And in the decoding stage only low-dimensional encoded information is used to reduce the number of channels, and the dimensions increase. (B) U-Net concatenation process. Specific location information is lost during dimensionality reduction in the encoding step, and that loss cannot be recovered in the decoding step. Therefore, a skip connection is used.
Fig 2.
Entire architecture of 3D U-Net for the RCT segmentation.
(A) Flowchart of the preprocessing and U-Net learning. The software selects T2 coronal data to be used for learning and loads labeling data. Both training and labeling data are converted into the Nifti image format. After learning with U-Net model, the result of prediction is generated in Nifti format. (B) Structure of U-Net model. The input data goes through the contracting path on the left and the expending path on the right. Both pathways are symmetrical to each other. Finally, it creates a segmentation map that classifies each pixel in the image.
Fig 3.
Segmented image of the rotator cuff tear region.
Each row is (A) ground truth based on manual labeling data by shoulder specialists, (B) raw MRI data, and (C) automatic segmentation results by the proposed learning model.
Fig 4.
Results of segmentation corresponding to rotator cuff tear site.
(A) Examples of the original MRI image which show rotator cuff tear. (B) The red area indicates the manually labeled region by shoulder specialists, and the blue area indicates the segmented region by the proposed learning model.
Fig 5.
Three-dimensional visualization of the rotator cuff tear region.
The final segmented image was converted into three-dimensional reconstructed image using in-house software. The maximal distance of tear was automatically measured and presented at the bottom of the image (in white letters) together with the tear classification.
Fig 6.
It shows automatically segmented RCT lesion (red area). Segmented area can be viewed two dimensionally in multiplanar (coronal, axial, sagittal) direction and controlled freely.
Fig 7.
Performance of the rotator cuff tear segmentation.
Dice, sensitivity, specificity, precision, F1 score, and Youden index.
Table 1.
Performance of the rotator cuff tear segmentation.
Values are presented as the median and interquartile range (Q1–Q3).
Fig 8.
Examples of incorrectly segmented cases.
In some cases of massive tears with cuff tear arthropathy, the result was interpreted as if there was no tear or a small-sized tear (A, B). In some cases of delamination tear patterns, the algorithm only recognized less retracted tear portions without detecting the more retracted medial ends (the red arrow) (C, D, and E).