E-TBNet: Light Deep Neural Network for Automatic Detection of Tuberculosis with X-ray DR Imaging

Abstract

Currently, the tuberculosis (TB) detection model based on chest X-ray images has the problem of excessive reliance on hardware computing resources, high equipment performance requirements, and being harder to deploy in low-cost personal computer and embedded devices. An efficient tuberculosis detection model is proposed to achieve accurate, efficient, and stable tuberculosis screening on devices with lower hardware levels. Due to the particularity of the chest X-ray images of TB patients, there are fewer labeled data, and the deep neural network model is difficult to fully train. We first analyzed the data distribution characteristics of two public TB datasets, and found that the two-stage tuberculosis identification (first divide, then classify) is insufficient. Secondly, according to the particularity of the detection image(s), the basic residual module was optimized and improved, and this is regarded as a crucial component of this article's network. Finally, an efficient attention mechanism was introduced, which was used to fuse the channel features. The network architecture was optimally designed and adjusted according to the correct and sufficient experimental content. In order to evaluate the performance of the network, it was compared with other lightweight networks under personal computer and Jetson Xavier embedded devices. The experimental results show that the recall rate and accuracy of the E-TBNet proposed in this paper are better than those of classic lightweight networks such as SqueezeNet and ShuffleNet, and it also has a shorter reasoning time. E-TBNet will be more advantageous to deploy on equipment with low levels of hardware.

Keywords: chest X-ray images; embedded device; neural network; tuberculosis detection.

Figure 1

Chest X-ray images of individual tuberculosis patients: (a) consolidation of the upper lobe of the right lung with cavities; (b) tuberculous exudative pleurisy; (c) secondary tuberculosis of right upper lung; (d) consolidation of right hilar with enlarged lymph nodes.

Figure 2

The residual block structure: (a) the traditional residual block structure; (b) the improved residual block structure. H, W, and C represent the height, width, and channel of the image, respectively; num represents the number of convolution kernels.

Figure 3

The framework of the proposed method of tuberculosis detection is based on chest X-ray images, consisting of two parts: the ECA block and the basic block. We found through a large number of ablation experiments that inserting the ECA block in this article after position 5 is the best. The model shown in Figure 3 is the optimal model.

Figure 4

Real image of the Jetson AGX Xavier device.

Figure 5

Examples of chest X-ray images from CHN datasets: (a) normal chest X-ray; (b) tuberculosis chest X-ray. Examples of chest X-ray images from MC datasets: (c) normal chest X-ray; (d) tuberculosis chest X-ray.

Figure 6

Visualization of KDE (kernel density estimation) distribution of two datasets.

Figure 7

The changes in model accuracy at training iterations.

Figure 8

Confusion matrices of CNNs in this paper on the testing set: (a) MobileNet_v3_small; (b) MobileNetV2; (c) Shufflenet_V2_x0_5; (d) ShuffleNet_V2_x1_0; (e) SqueezeNet1_1; (f) E-TBNet.

Figure 8

Confusion matrices of CNNs in this paper on the testing set: (a) MobileNet_v3_small; (b) MobileNetV2; (c) Shufflenet_V2_x0_5; (d) ShuffleNet_V2_x1_0; (e) SqueezeNet1_1; (f) E-TBNet.

Figure 9

Comparison of the reasoning time of the CNNs used in this paper.

Figure 10

The results of five classification algorithms under different characteristics.