Efficient Deep Network Architectures for Fast Chest X-Ray Tuberculosis Screening and Visualization

Abstract

Automated diagnosis of tuberculosis (TB) from chest X-Rays (CXR) has been tackled with either hand-crafted algorithms or machine learning approaches such as support vector machines (SVMs) and convolutional neural networks (CNNs). Most deep neural network applied to the task of tuberculosis diagnosis have been adapted from natural image classification. These models have a large number of parameters as well as high hardware requirements, which makes them prone to overfitting and harder to deploy in mobile settings. We propose a simple convolutional neural network optimized for the problem which is faster and more efficient than previous models but preserves their accuracy. Moreover, the visualization capabilities of CNNs have not been fully investigated. We test saliency maps and grad-CAMs as tuberculosis visualization methods, and discuss them from a radiological perspective.

Figure 1

Schematic representation of the network architecture. conv = convolution, pool = pooling, GAP = Global Average Pooling, fc = fully connected. Convolutions and pooling sizes are reported as height × width/stride. The additional number indicates the number of feature maps for convolutions and the number of output neurons for the fully connected layer. Circled pluses indicate an addition operation.

Figure 2

Receiver Operation Characteristic (ROC) curves of the 5-fold cross-validation study on three different datasets. The accuracy and Area Under the ROC (AUC) score are (a) 0.811 for the Montgomery dataset, (b) 0.9 for the Shenzhen dataset and (c) 0.925 for the combined dataset. For a description of the three datasets refer to Table 1.

Figure 3

Saliency map with overlay for two correctly classified cases. Panels (a) and (d) show the chest images of the patients, panels (c) and (f) show the saliency maps, while panels (b) and (e) show the saliency maps overlaid on the chest images for comparison. The first row shows a patient with tuberculosis, with output score 0.98 (the maximum was 1). The second row shows a healthy patient with score 0.00 (the minimum was 0). Both scores suggest high confidence in the prediction.

Figure 4

Saliency map with overlay for two misclassified patients. Panels (a) and (d) show the chest images of the patients, panels (c) and (f) show the saliency maps, while panels (b) and (e) show the saliency maps overlaid on the chest images for comparison. The first row shows a healthy patient classified with output score 0.98 (the maximum was 1). The second row shows a patient with tuberculosis with output score 0.04 (the minimum was 0). Both scores suggest serious misclassifications.

Figure 5

Gradient class activation maps (grad-CAMs) for different layers of the same patient. Panel (a) shows layer 1, (b) layer 4, (c) layer 7, (d) layer 10 and (e) layer 13. The scales of each panel are independent of each other. The patient is also shown in the first row of Fig. 3 (the true positive). The activation maps are calculated for the positive class and are shown for the last layer of each convolutional block (e.g. the one just before pooling). Activation maps of the higher layers show higher level features, which should be tuberculosis specific.

Figure 6

Saliency map of our network trained on the combined dataset, using about 660 patients for training. Panels (a) shows the chest images of the patients, panel (c) shows the saliency maps, while panel (b) shows the saliency maps overlaid on the chest images for comparison. This figure should be compared to the first row of Fig., which shows the saliency map predicted for the same patient, but with the network trained only on the Montgomery dataset. The localization ability of this saliency map is drastically improved. The image appears less noisy in unimportant regions and more intense in the areas where tuberculosis is really present.