Deep Convolutional Neural Networks for Endotracheal Tube Position and X-ray Image Classification: Challenges and Opportunities

Abstract

The goal of this study is to evaluate the efficacy of deep convolutional neural networks (DCNNs) in differentiating subtle, intermediate, and more obvious image differences in radiography. Three different datasets were created, which included presence/absence of the endotracheal (ET) tube (n = 300), low/normal position of the ET tube (n = 300), and chest/abdominal radiographs (n = 120). The datasets were split into training, validation, and test. Both untrained and pre-trained deep neural networks were employed, including AlexNet and GoogLeNet classifiers, using the Caffe framework. Data augmentation was performed for the presence/absence and low/normal ET tube datasets. Receiver operating characteristic (ROC), area under the curves (AUC), and 95% confidence intervals were calculated. Statistical differences of the AUCs were determined using a non-parametric approach. The pre-trained AlexNet and GoogLeNet classifiers had perfect accuracy (AUC 1.00) in differentiating chest vs. abdominal radiographs, using only 45 training cases. For more difficult datasets, including the presence/absence and low/normal position endotracheal tubes, more training cases, pre-trained networks, and data-augmentation approaches were helpful to increase accuracy. The best-performing network for classifying presence vs. absence of an ET tube was still very accurate with an AUC of 0.99. However, for the most difficult dataset, such as low vs. normal position of the endotracheal tube, DCNNs did not perform as well, but achieved a reasonable AUC of 0.81.

Keywords: Artificial intelligence; Artificial neural networks (ANNs); Classification; Machine learning; Radiography.

Figures

Fig. 1

ROC curves of the best-performing classifiers for the three different datasets. The AUC of the chest/abdomen model was the highest at 1.000 (solid black line), followed by the presence/absence at 0.989 (dashed black line), and low/normal at 0.809 (dotted black line)

Fig. 2

With reduced learning rate of 0.001, there is fast convergence using pre-trained weights on the AlexNet model. Both training and validation loss reduce over the course of training, with accuracy on the validation dataset reaching approximately 99% at 90 epochs

Fig. 3

With higher learning rate of 0.01, there is no convergence after 90 epochs using pre-trained weights on the AlexNet model, likely due to large gradients that impact the learning process. The training and validation loss remains high with accuracy constant at 50%

Fig. 4

Effect of number of training cases on AUCs for the pre-trained models for the ET presence/absence dataset. Using 100% of training cases resulted in statistically significant higher AUCs than with 25% of the data (P = 0.015 and P < 0.0001, for AlexNet_T and GoogLeNet_T, respectively)

Fig. 5

Chest radiograph with ET tube present (white arrow)

Fig. 6

Follow-up chest radiograph on the same patient after ET tube has been removed. The white arrow depicts the absence of the ET tube

Fig. 7

Cropped chest radiograph image centered on the carina. The ET tube tip is low at the orifice of the right main stem bronchus (white arrow)

Fig. 8

Cropped chest radiograph image shows that the ET tube tip (white arrow) is in satisfactory position 4 cm above the carina

Fig. 9

Saliency map on the left with corresponding radiograph on the right, where an ET tube is present. The saliency map highlights parts of the images that contribute most to the prediction class. The black arrows points to the location of the ET tube, which has a visible contribution to the prediction label. However, there are other parts of the image, including edges of ribs and the aortic knob that also contribute to the score for this image. While the network had high accuracy, there is still room for improvement

Fig. 10

Saliency map on the left with corresponding radiograph on the right with low position of ET tube in the right main stem bronchus. On the saliency map, the black arrows point to the location of the ET tube. The white arrows point to an unrelated overlying catheter and an enteric tube. From the saliency map, the ET tube, enteric tube, and overlying catheter all have contributions to the prediction class, indicating that the network is not as well formed and is inferring from parts of the image (enteric tube and overlying catheter) that are not relevant to the prediction label (ET tube is low). This may explain the lower accuracy of this network compared to the others