Keywords

1 Introduction

Work-related musculoskeletal disorder (WMSDs) refers to systemic muscle, bone, and nervous system disorders caused by occupational factors. During the work process, the musculoskeletal system needs to bear the ergonomic load such as posture load and strength load, because the operator needs to maintain repetitive movements or a compulsory position to carry or lift heavy objects. And, the musculoskeletal system of the limbs is damaged, and there are irritations such as acid, numbness, swelling, and pain [1]. More than 1.7 billion people worldwide are affected by musculoskeletal diseases, so it is important for the abnormal detection of musculoskeletal.

In recent years, deep learning has become a hot spot in the field of machine learning research. Image features extracted by deep convolutional neural networks (DCNN) have proven to be effective in image classification, segmentation, or retrieval applications. Based on its advantages in non-medical images, DCNN is beginning to be gradually applied to medical image classification and detection problems. For example, Spanhol proposed the use of DCNN to classify breast cancer pathology images [2]. Li proposed a DCNN-based pulmonary nodule classification system [3]; Roth proposed an algorithm for developing a lymph node detection system using DCNN [4]. Based on the above successful experience of DCNN applied to medical images, we tried to use DCNN to classify radioactive musculoskeletal. However, the images currently used to train the DCNN model have the following drawbacks:

  1. 1.

    The quality of the training images is not high.

  2. 2.

    For radioactive image datasets, DCNN does not extract its texture features very well.

Therefore, this paper proposes to preprocess the image to improve the image quality, then fuse the shallow texture features and then send them into the Densnet169 network for classification, and compare the classification effects without using the shallow texture features.

2 Musculoskeletal Abnormality Diagnosis Based on Densenet

The method proposed in this paper is shown in Fig. 1 including image preprocessing, image fusion, and training depth neural network.

Fig. 1.
figure 1

The structure of the algorithm

Full size image

2.1 Preprocessing

Histogram equalization is a method of enhancing image contrast (image contrast). In experiments, we used contrast limited adaptive histogram equalization (CLAHE) [5] for image enhancement. The histogram value of the CLAHE method is:

$$ \text{Hist}^{{\prime }} (i) = \left\{ {\begin{array}{*{20}l} {\text{Hist}(i) + L} \hfill & {\text{Hist}(i) < T} \hfill \\ {H_{\max} } \hfill & {\text{Hist}(i) \ge T} \hfill \\ \end{array} } \right. $$

where \( {\text{Hist}}(i) \) is the derivative of the cumulative distribution function of the sliding window local histogram, and \( H_{ \hbox{max} } \) is the maximum height of the histogram. We cut off the histogram from the threshold T and then evenly distribute the truncated portion over the entire grayscale range to ensure that the total histogram area is constant, so that the entire histogram rises by a height L. The preprocessed image is shown in Fig. 2.

Fig. 2.
figure 2

The original image on the left and the preprocessed image on the right

Full size image

2.2 DenseNet Neural Networks

In the field of imagery, deep learning has become a mainstream method. We chose DenseNet [6] as our model because it has better performance than ResNet in the case of less parameter and computational cost.

DenseNet’s structure is to interconnect all layers, specifically each layer will accept all of its previous layers as its additional input, as shown in Fig. 3, and DenseNet directly merges feature maps from different layers, which can be achieved feature reuse to improve efficiency. We experimented with the DenseNet 169 layer network structure, adjusting the number of neurons in the last layer to 2.

Fig. 3.
figure 3

DenseNet structure

Full size image

2.3 Merge Texture Features

Deep neural network can directly classify images, but it does not use the underlying features of images. Considering that there are a large number of underlying features such as texture in medical images, this paper integrates the original image and texture features and then sends them into the deep neural network for classification.

2.3.1 Extract the LBP Texture Features of the Image

The local binary mode is a texture metric in the gray range. In order to improve the limitations of the original LBP, it is impossible to extract the texture features of large-size structures. Ojala et al. [7] modified the LBP to obtain the LBP value that uniquely represents the local texture features:

$$ {\text{LBP}}_{P,R} = \mathop \sum \limits_{i = 0}^{P - 1} s \left( {g_{i} - g_{c} } \right)2^{i} $$
(1)

where \( {\text{S}}( \cdot ) \) is defined as:

$$ {\text{s(x)}} = \left\{ {\begin{array}{*{20}l} {1,} \hfill & {x \ge 0} \hfill \\ {0,} \hfill & {x < 0} \hfill \\ \end{array} } \right. $$
(2)

2.3.2 Merging LBP Features

Performing mean variance normalization on the grayscale image and the corresponding LBP feature image, and then combining the grayscale image with the LBP image in the channel dimension, for example, the shape of the grayscale image is (h, w), the shape of the LBP image Also (h, w), the combined shape is (h, w, 2).

3 The Experiment

3.1 Experimental Dataset

MURA is a dataset of musculoskeletal radiographs, which contains a total of 14,863 studies of 12,173 patients and 40,561 multiview radiographs. Each was one of seven types of standard upper extremity radiology studies: fingers, elbows, forearms, hands, humerus, shoulders, and wrists [8].

MURA contains 9045 normal and 5818 abnormal musculoskeletal radiographic studies as shown in Table 1.

Table 1. Composition of the MURA dataset
Full size table

3.2 The Experiment

We trained a 169 layer convolutional neural network to diagnose musculoskeletal abnormalities.

3.2.1 Parameter Settings

During training, we resize the image to 320 × 320, and we use a random horizontal flip and a random rotation of 30° for augmentation. All parameters of the network are randomly initialized. The optimizer chooses Adam, where \( \beta_{1} \) is equal to 0.9 and \( \beta_{2} \) is equal to 0.999. We trained the model using minibatches of size 16. The initial learning rate is chosen to be 0.00625. When iterating to the 60th and 80th, the learning rate is attenuated 10 times.

3.2.2 The Experimental Setup

To demonstrate the effectiveness of our proposed method, we conducted three sets of experiments:

  • Experiment 1: Training with grayscale image

  • Experiment 2: Training with grayscale + lbp (radius 1, sampling 8 points)

  • Experiment 3: Training with grayscale + lbp (radius 2, sampling 8 points).

3.3 The Experimental Results

The performance of the experimental results on the validation set is shown in Fig. 4. We can see that combining LBP features when the model converges is better than using only grayscale images, which proves the effectiveness of the proposed fusion method between gray images and shallow texture features.

Fig. 4.
figure 4

The performance of the three experiments on the validation set

Full size image

Table 2 shows the highest precision achieved in the validation set for Experiment 1, Experiment 2, and Experiment 3.

Table 2. The highest precision achieved by the three experiments on the validation set
Full size table

4 Conclusion

In this paper, a new classification algorithm based on deep convolutional neural network is proposed for radioactive musculoskeletal images. Adaptive histogram equalization with limited contrast was used to improve the training image quality, and the image texture features were merged and sent to DenseNet169 for training and classification. Experimental analysis shows that the proposed algorithm can effectively improve the classification accuracy of radioactive musculoskeletal images.

Further work in this paper is to continue to optimize the algorithm and expand the dataset to train a new depth neural network sensitive to radioactive musculoskeletal images.