Medina et al., 2020. Deep learning method for segmentation of rotator cuff muscles on MR images

# 세줄 요약 #

1. To develop and validate a deep convolutional neural network (CNN) method capable of (1) selecting a specific shoulder sagittal MR image (Y-view) and (2) automatically segmenting rotator cuff (RC) muscles on a Y-view.
2. For model A, we manually selected shoulder sagittal T1 Y-view from 258 cases as ground truth to train a classification, For model B, we manually segmented subscapularis, supraspinatus, and infraspinatus/teres minor on 1048 sagittal T1 Y-views.
3. A showed top-3 accuracy > 98% to select an appropriate Y-view and accurate RC muscle segmentations with mean Dice scores > 0.93.

 

 

# 상세 리뷰 #

1. introduction

• Rotator Cuff (RC) tendon tears are associated with varied degrees of muscle atrophy, manifested by decreased muscle bulk and fatty infiltration.
• MRI is the reference standard for imaging RC tendons for tears, severity of cuff abnormalities, and postoperative healing.
• We need accurate and automated segmentation strategies for estimation of RC muscle atrophy.

Fig 1. (a) Definition of sagittal Y-view, (b) Grouping of sagittal images in 3 anatomic zones [Ref. Medina et al., 2020]

 

2. Method

• The purpose of our study was to develop deep convolutional neural networks (CNN) to identify a scapular Y-view.
• Our study was IRB-approved Dataset
  • internal dataset
    • Institution: General Electric, Waukesha, WI, USA; Hitachi Medical Corporation, Tokyo, Japan; Siemens Healthcare, Erlangen, Germany; Philips Healthcare, Amsterdam, Netherlands
    • Duration: 2018.10 ~ 2020.01
  • external dataset
    • non-affiliated imaging facilities that were uploaded to our hospital's database.
• We created two models.
  • Model A. Y-view selection = classification CNN
    • Inception V3 (16 batch, 100 epochs, droput 0.2, learning rate 0.001, optimizer RMSprop)
    • train dataset: 258 cases, test dataset: 100 internal, 50 external cases.
  • Model B. RC muscle segmentation = segmentation CNN
    • modified U-Net (8 batch, 50 epochs, dropout 0.25, learning rate 0.0001, softmax)
    • train dataset: 1048 images, test dataset: 105 internal, 50 external images.

Fig 2. Workflow for testing of model A. Ground truth (zone 1. most lateral image to lateral acromioclavicular (AC) joint, zone 2. AC joint up to Y-view, zone 3. Y-view to most medical images) [Ref. Medina et al., 2020]

 

3. Result

• Model A (Y-view selection). top-3 accuracy 98%
  • internal test cases: 98.7 ± 1.0%
  • external test cases: 99.7 ± 1.0%

• Model B (RC muscle segmentation). mean Dice scores > 0.93

Tab 1. Mean Dice, Precision, Recall scores for model B. [Ref. Medina et al., 2020]

Fig 3. Examples of accurate muscle segmentation using model B. [Ref. Medina et al., 2020]

 

* Reference: Medina, Giovanna, et al. "Deep learning method for segmentation of rotator cuff muscles on MR images." Skeletal Radiology 50.4 (2021): 683-692.