Shim et al., 2020, Automated rotator cuff tear classification using 3D convolutional neural network

# 세줄 요약 #

1. Rotator cuff tear (RCT) is one of the most common shoulder injuries. When diagnosing RCT, skilled orthopedists visually interpret magnetic resonance imaging (MRI) scan data.
2. MRI data from 2,124 patients were used to train and test the VRN-based 3D CNN to classify RCT into five classes (None, Partial, Small, Medium, Large-to-Massive).
3. The VRN-based 3D CNN outperformed orthopedists specialized in shoulder orthopedists in binary accuracy (92.5 % vs. 76.4%), top-1 accuracy (69.0% vs. 45.8%), top-1±1 accuracy (87.5% vs. 79.8%).

 

# 상세 리뷰 #

* Author: Eungjune Shim, Joon Yub Kim, Jong Pil Yoon, Se‑Young Ki, Taewoo Lho, Youngjun Kim & Seok Won Chung

 

1. Introduction

• Convolutional neural networks (CNN) have frequently been shown to be successful at medical image classification.
  • For example, Inception has shown specialist-level performance in classifying skin cancer, shoulder fracture detection, diabetic retinopathy.
• Much of the existing research on diagnosis using the classification method have limitations.
  • 1. The preprocessing learning data requires the time-consuming manual labor of clinical experts
    • because well-trained CNN need precise localization labels or segmentation masks from raw images.
  • 2. Most of the previous methods of CNN(Vision deep learning) algorithms have focused on only 2D-based CNNs.
    • Those 2D-based CNN outputs are the number of slices diagnosis, so the diagnosis result must be derived through some process that combines the different results into a single outcome.
    • 2D slice-based CNN methods are susceptible to missing meaningful features found in 3D volume data.
• To overcome these problems, we propose a deep 3D-CNN method to classify rotator cuff tear (RCT), one of the most common shoulder injuries.
  • The diagnosis of an RCT is largely based on the patient's symptoms, and accurate diagnosis is carried out using medical images
    • MRI data are currently widely used to diagnose RCTs.
  • Only the patient's diagnosis information (normal or RCT size) and simple region of interest (ROI) selection are needed for preprocessing.
  • Our method uses fairly deep 3D CNN
    • three-dimensionally extended inception-ResNet = Voxception-ResNet = VRN
    • we also propose an effective visualization method using 3D volume rendering of class activation map (CAM).

Voxception-ResNet Architecture [https://slideplayer.com/slide/14678795/]

• As a result, the proposed method automatically determines the RCT size in five categories and visualizes the 3D localization information of RCT as a second opinion for clinical decisions.
  • 5 categories: None, Partial, Small, Medium, Large-to-Massive

Fig 1. The network structure of the proposed method for automated rotator cuff tear (RCT) diagnosis. Original shoulder MRI volume is cropped and sampled into 64 × 64 × 64 volume. The sampled volume is feed-forwarded to the VRN-based 3D CNN to classify the RCT case (None, Partial, Small, Medium, and Large-to-Massive). [Shim et al., 2020, Sci Rep]

 

 

2. Result

• Experiments
  • Dataset: Total 2,124 MRI datasets (200 Test sets, 1,924 Training sets)
  • Model: VRN-based 3D CNN (All training were conducted until epoch 100)

Table 1. MRI data sets used for training and test. [Shim et al., 2020, Sci Rep]

• Evaluation of the 3D CNN algorithm
  • 1. Diagnosing RCTs (None vs RCT = binary classification): binary accuracy, sensitivity, specificity, precision and F1-score.
  • 2. Classifying RCTs (5 categories of RCT): top-1 accuracy, top-1±1 accuracy, sensitivity, specificity and diagnostic time.
    • top-1±1 accuracy: Regarding one-size prediction error as correct (ex. Predicted tear "medium" = Correct in ground truth: "Small", "Medium", "Large")
• Evaluation of the diagnostic performance of human readers
  • Comparing the performance of diagnosis and classification of RCT of 3D CNN with those of human readers
    • Human readers: 13 general orthopedists & 4 orthopedic shoulder specialists
    • Comparing Scores (CNN vs. Human): the top-1 accuracy, top-1±1 accuracy, sensitivity, specificity, and diagnostic time

Fig 2. A screen shot of the developed software for automated diagnosis and 3D visualization of RCT by the VRN-based 3D CNN. [Shim et al., 2020, Sci Rep]

• Statistic 1.  Diagnosing RCTs (None vs. RCT): binary accuracy
  • VRN-based CNN: 92.5%
  • orthopedic shoulder specialists: 76.4% 
  • general orthopedists: 68.2%

Table 2. RCT diagnosis performance of 3D CNN for binary classification. [Shim et al., 2020, Sci Rep]

Fig 3. Comparision of diagnosis performance between clinical experts and the proposed method. (A) ROC curve, (B) Binary accuracy. [Shim et al., 2020, Sci Rep]

• Statistic 2. Classifying RCTs (5 categories of RCT): top-1±1 accuracy
  • VRN-based CNN: 87.5%
  • orthopedic shoulder specialists: 79.8% 
  • general orthopedists: 71.0%Statistics

Table 3. Comarision of the performance for multi-class RCT classification between groups. [Shim et al., 2020, Sci Rep]

Fig 4. Comparision of diagnosis performace between clinical experts and the proposed method. (A) Top-1 accuracy; (B) diagnostic time. [Shim et al., 2020, Sci Rep]

* Reference: Shim, E., Kim, J.Y., Yoon, J.P. et al. Automated rotator cuff tear classification using 3D convolutional neural network. Sci Rep 10,15632 (2020). https://doi.org/10.1038/s41598-020-72357-0