Shoulder Bone Segmentation with DeepLab and U-Net

Abstract

Evaluation of 3D bone morphology of the glenohumeral joint is necessary for pre-surgical planning. Zero echo time (ZTE) magnetic resonance imaging (MRI) provides excellent bone contrast and can potentially be used in place of computed tomography. Segmentation of shoulder anatomy, particularly humeral head and acetabulum, is needed for detailed assessment of each anatomy and for pre-surgical preparation. In this study we compared performance of two popular deep learning models based on Google's DeepLab and U-Net to perform automated segmentation on ZTE MRI of human shoulders. Axial ZTE images of normal shoulders (n=31) acquired at 3-Tesla were annotated for training with a DeepLab and 2D U-Net, and the trained model was validated with testing data (n=13). While both models showed visually satisfactory results for segmenting the humeral bone, U-Net slightly over-estimated while DeepLab under-estimated the segmented area compared to the ground truth. Testing accuracy quantified by Dice score was significantly higher (p<0.05) for U-Net (88%) than DeepLab (81%) for the humeral segmentation. We have also implemented the U-Net model onto an MRI console for a push-button DL segmentation processing. Although this is an early work with limitations, our approach has the potential to improve shoulder MR evaluation hindered by manual post-processing and may provide clinical benefit for quickly visualizing bones of the glenohumeral joint.

Keywords: DeepLab; MRI; U-Net; ZTE; glenohumeral; glenoid; humeral head; image processing.

Figure 1.

(A) Anatomy of shoulder showing major bone structures of humerus, humeral head, glenoid, scapula. Axial imaging plane is shown in dotted red box. Conventional shoulder MR images acquired using conventional (B) spin echo proton density weighted, (C) spin echo proton density weighted with fat suppression, and (D) zero echo time (ZTE) sequences. Conventional sequences do not isolate bone effectively. (E) Inverted ZTE image depicting mostly bony tissues with high signal intensity. Conventional MR images depict non-bony tissues with similar contrast as bony tissues, making them less useful for bone-only imaging.

Figure 2.

Inverted axial ZTE shoulder images used in this study were acquired with moderately varying scan parameters. (A) was acquired with TR=88 ms, TE=0.016 ms, FOV=180 mm, matrix=256×256, and 1 mm slice thickness. (B) was acquired with TR=458 ms, TE=0.016 ms, FOV=160 mm, matrix=256×256 and 1 mm slice thickness. (C) was similar to (B) but acquired with FOV=180 mm, matrix=512×512, and 1.2 mm slice thickness. While varying in image contrast, all images shared a similar feature of depicting bones of the shoulder with high signal intensity.

Figure 3.

Manual segmentation of the MRI images. (A) Inverted ZTE MRI shoulder images acquired in the axial plane were manually annotated (segmented) into (B) background, (C) humeral head / humerus, and (D) the remaining tissues.

Figure 4.

Architectures of (A) U-Net and (B) DeepLab used in this study. Adapted from [22] and [37], respectively. Training results showing accuracy and loss values for (C) U-Net and (D) DeepLab.

Figure 5.

(A) Flow chart of ZTE DL processing, which reads raw DICOM images, performs DL segmentation to create masks for humerus and the remaining tissues. The masks are then multiplied with the raw image to create segmented DICOM images that are saved as new series in the exam. (B) Segmented DICOM images viewed in a PACS viewer, showing the original image on the left, segmented humeral bone in the middle, and segmented remaining tissues on the right.

Figure 6.

Segmentation results on test images. (A, D) Ground truth or manually segmented images of humeral bone and the remaining other tissues shown for comparison. Output segmented images of (B, C) the humeral head (E, F) and the remaining tissues after DL segmentation performed by (B,E) U-Net and (C,F) DeepLab. Qualitatively, U-Net slightly over-estimated area for humeral head while DeepLab slightly under-estimated. (G) Input ZTE MRI image is shown. (H, I) Segmented ZTE images (from the ground truth; A and D) were used to create separate 3D renderings of the (H) humerus and (I) glenoid / scapular bone.

Figure 7.

DL model performances compared. Boxplots of inference accuracy (Dice score, sensitivity, specificity) quantified on the humeral bone (A) and the remaining tissue (B), determined using U-Net (blue) and DeepLab (red) models. Marked differences in the accuracy metrics for the humeral bone was noted.

Figure 8.

Comparison of MRI vs. CT segmentation. ZTE MRI (A) and CT (B) data of the same subject were registered and segmented (using U-Net for MRI, manually for CT). The segmented images were fused (C), showing the overlapping regions as white, and the non-overlapping regions in magenta for MRI and green for CT.