Chronic lung disease (CLD) is the most common and serious lung lesion in extremely preterm infants. Early intervention in predicted CLD patients is effective in improving prognosis. However, methods for early detection of onset have yet to be established. Previous studies have utilized clinical data to predict the CLD, but none have employed chest X-ray images. This paper proposes a novel method for predicting CLD onset on chest X-rays using convolutional neural networks (CNNs). We also compare the performance of CLD prediction among 10 state-of-the-arts CNNs, and conducted a longitudinal study to assess change in prediction accuracy at each of 3, 7, 14, and 28 days of age. Validation was conducted with 115 cases (51 CLD and 64 normal). The proposed method achieved the best performance with an accuracy of 0.715, 0.652, 0.739, 0.765, and an AUC of 0.796, 0.673, 0.778, 0.828 at 3, 7, 14, 28 days, respectively. The CNN selected was NASNet_Large, DenseNet121, DenseNet169, DenseNet201, at 3, 7, 14, 28 days, respectively. Our findings conclude that the performance of image-based prediction is comparable to the previous study using clinical data, and further improvement can be achieved by incorporating both clinical data and images.
