Prediction of the duration needed to achieve culture negativity in patients with active pulmonary tuberculosis using convolutional neural networks and chest radiography.

BACKGROUND

We aimed to predict the duration needed to achieve culture negativity in patients with active pulmonary tuberculosis using convolutional neural networks (CNNs) and chest radiography.

METHODS

Medical records were searched for eligible patients with culture-confirmed active pulmonary tuberculosis. The eligible patients were randomly assigned to the training dataset group (N = 180) and the validation dataset group (N = 59). Posteroanterior X-ray radiographs in the standing position were obtained at diagnosis. The image data were augmented by a factor of 10 by randomly shifting and rotating the original image. Thus, 1800 images (112 × 112 pixels, 8-bit grayscale) from 180 patients in the training dataset group were used for training the CNN model. The model performance was evaluated on the validation dataset.

RESULTS

The values predicted by the CNN model were significantly associated with the actual values (Pearson's correlation coefficient 0.392, p = 0.002). The mean absolute error was 18.0. The visualization of the layer outputs suggested that the CNN model recognized some of the chest radiographic findings that were useful in predicting the duration needed to achieve culture negativity.

CONCLUSIONS

The CNN model was useful for predicting the duration needed to achieve culture negativity in active pulmonary tuberculosis, although the accuracy was unsatisfactory. This study suggests that chest radiography findings are as important as other clinical factors for prediction and could be learned by the machine.