2.1

Structure of network

At present, the more advanced one-stage object detection network is the FCOS network proposed by Tian et al.¹², it has a faster detection speed than the two-stage detection network and a similar detection accuracy, which is the basis of this paper.

FCOS network achieves proposal-free and anchor-free by many clever designs (FPN, center-ness, scale limitation of feature point regression in each layer, etc.). At the same time, it avoids complex IOU calculation and the matching between anchor and GT Bounding box during training. The network architecture of this study can be seen in Figure 1, which adopts feature pyramid Network (FPN) and three-branch head detection network.

Figure 1.

Structure of the network.

2.2

Self-calibrated convolutions

In order to further optimize the network parameters on the foundation of the FCOS network and improve network accuracy, this paper draws on the SC-Net suggested by Liu et al.¹³. The network mainly uses a self-calibrating convolutional SCConv, whose structure is shown in Figure 2.

Figure 2.

Structure of the SCConv module.

The design of SCConv is simple and versatile, and can easily enhance the performance of standard convolution layers without introducing additional parameters and complexity. The architecture of the SC module is shown in Figure 3.

Figure 3.

Schematic illustration of the self-calibrated convolutions.

2.3

Squeeze-and-excitation blocks

To better exploit the dynamic relationship between feature channels, this paper introduces the SE-Net proposed by Hu et al.¹⁴. SE-Net is very simple in construction and apt to instruct, there is no need to introduce new functions or layers, with good performance in the aspect of parameter complexity and network structure. The network structure is shown in Figure 4. Figure 5 reveals the structure of embedding the SE module into the Res-Net module.

Figure 4.

Squeeze-and-excitation block.

Figure 5.

The architecture of the SE-ResNet module.

In this paper, we use SE-ResNet to focus on the channel relationship of the network, Squeeze operation to establish the dependency relationship between channels, and Excitation operation to recalibrate the features. The combination of the two emphasizes the useful features and suppresses the useless features, which can effectively improve the model performance and increase the accuracy rate.

2.4

Soft-NMS

In our study, we use Soft-NMS¹⁵, which mainly solves the problem of excessive deletion of boxes by NMS, Soft-NMS has learned the lessons of NMS, during the execution of the algorithm, instead of simply deleting the detection box whose IOU is above the threshold, lower its score. The algorithm process is the same as NMS, but the function operation is used for the original confidence score, and the goal is to reduce the confidence score. Soft-NMS is expressed as follows.

2.5.

Improved res module

Based on the improvement of the traditional Res-Net network, the SC module and SE module are introduced to modify the residual module in the traditional Res-Net network. The improved module architecture is shown in Figure 6.

Figure 6.

The architecture of the improved Res-Net module.

2.6

Loss function

Three parts constitute the loss function of this network, Focal loss for classification loss, IOU loss for regression loss and center-ness loss for BCE. The SoftMax is discarded in the classification loss, and the sigmoid function is used for each channel (each channel represents a category) of the classification output by the head, and then the Focal loss is used. IOU Loss only performs regression calculation for those meaningful feature points. The specific loss function is as below.

3.1

Implementation details

The experimental environment of this paper is 20.04.3-Ubuntu system, the processor model is Intel(R) Xeon(R) CPU E5-2640 v4 @ 2.40GHz, the graphics card model is Nvidia Quadro M4000 8GB, the python version used is 3.6.13. We use pytorch to build the deep learning framework.

In this paper, we use a non-public human lumbar disc MRI-T2 image dataset collected from the Internet with 470 images, which is divided into train dataset (376 images), val dataset (47 images) and test dataset (47 images) by 8:1:1. The target categories in the dataset are divided into normal and herniated. The training process is optimized using stochastic gradient descent (SGD) with an initial learning rate of 0.005 and momentum of 0.9.

3.2

Results on MRI

The experiment was carried out according to the five-fold cross-validation, take the mean value of five experimental results as the final experimental result. Partial detection results of the model on MRI dataset are shown in Figure 7.

Figure 7.

Some detection results on spine MRI dataset.

Laboratory result of some model on MRI test set are listed in Table 1. The network model proposed in this paper has few parameters, and its detection speed is faster and the detection accuracy is the highest among all models. Compared with the generic Faster R-CNN and Retina-Net, it has a large performance improvement, which fully reflects the practical application value of the model. Moreover, our network has no complex network structure and has great advantages in the deployment of the network model.

Table 1.

Comparisons on spine MRI dataset.

3.3

Ablation study

For verifying the availability of the modules introduced in our network, we conducted ablation experiments on the main modules to check on their role in the network. The results of ablation experiments are shown in Table 2. We can see that the module we introduced improves the performance of the network.

Table 2.

Ablation study on spine MRI dataset.

Furthermore, we performed an ablation experiment on the SE modules to evaluate the effect of their alignment position when integrating them into existing frameworks. Figure 8 shows the structure of these variants, and Table 3 shows the property of these variants. As can be seen from the laboratory results, the performance of SE-PRE block, SE-Identity block, and Standard SE block is similar, while the use of the SE-POST block results in a degradation in capability. The test shows that the property improvement brought by the SE block is robust to its position change if it is used before branch aggregation.

Figure 8.

SE block with different structures.

Table 3.

Effect of different SE block.