Current and emerging artificial intelligence applications in chest imaging: a pediatric perspective

Automated lung nodule detection is the most studied application for chest radiography. Research that investigated the performance of lung nodule detection in chest radiographs started decades ago. With few publicly available datasets, systems were hard to compare. Steady improvement has been shown on the Japanese Society of Radiological Technology (JSRT) dataset consisting of chest radiographs with and without lung cancer, which can be used for validation purposes. First studies from 1999 reached a sensitivity of 35% at 6 false positives per image [30]; by 2014, this improved to a sensitivity of 75% at 0.5 false positives per image on this dataset [31]. Recently developed AI systems that incorporated deep learning have not tested their performances on the JSRT data but do report improved performances up to a sensitivity of about 80% at only 0.05 false positives per image [4].

For lung nodule detection, performance of these AI systems is similar or even surpasses the performance of radiologists. In a study from Nam et al. [4], a deep learning system that was trained on more than 40,000 chest radiographs showed in receiver operating characteristics (ROC) analysis an area under the curve (AUC) of 0.92–0.99 (depending on the validation set) for the detection of lung nodules with an average size ranging 20–40 mm. AI performance was similar or significantly better than that of the participating thoracic radiologists in this study. Most readers had improved performance when aided by the deep learning system.

Another recent study, by Yoo et al. [32], assessed the performance of their algorithm on chest radiographic screening data from the National Lung Screening Trial. On the baseline screening data, consisting of 5,485 radiographs, the algorithm reached a sensitivity of 94.1% and a specificity of 83.3% for detecting malignant pulmonary nodules.

Up to now, 11 certified products have been made available to help detect pulmonary nodules in chest radiographs. The developed systems have been trained on adult data, often older people with a high risk of lung cancer. However, the spectrum of abnormalities in the pediatric chest is different from that in adults. In the pediatric population, it would be more helpful to detect pulmonary metastases instead of potential lung cancers, which is relevant, for example, in children with osteosarcoma. However, such systems that aim to detect pulmonary metastases specifically are not yet available, and current systems trained on lung cancer data probably have a lower performance for the detection of pulmonary metastases. Furthermore, the pediatric chest is quite different from the adult chest with respect to shape, mediastinal contours and density of bone structures, which also might lower the performance of the current systems when applied to this population.

In the last few years, several studies have been published on the detection and diagnosis of consolidations in chest radiographs. Several groups have worked on AI to detect pulmonary tuberculosis in chest radiographs because resources and trained personal in developing countries are scarce, so automated evaluation of chest radiographs is especially beneficial in these countries. One of the first algorithms for tuberculosis, developed by Hogeweg et al. [33], showed a slightly lower performance than the observer for the detection of tuberculosis in a high-incidence population. Qin et al. [10] compared three available commercial deep learning systems and found similar performance of all systems for the detection of tuberculosis in chest radiographs, with an AUC ranging from 0.92 to 0.94 as compared with molecular testing. The AI systems reached higher specificity than the participating radiologists in the study, at a matched sensitivity level [10].

In addition to studies on AI for tuberculosis, studies are performed for AI that can help detect pneumonia [34‐36]. In one of these studies [34], CheXNeXt, a convolutional neural network developed by the Stanford Machine Learning Group that can detect 14 pathologies on anteroposterior (AP) or posteroanterior (PA) chest radiographs, was tested on the chestX-ray8 dataset [37]. For 11 of the 14 pathologies, CheXNeXt achieved radiologist-level performance, including consolidations (AUC of 0.84 for radiologists versus 0.89 for the algorithm) [34]. In this study, radiologists achieved higher performance for the detection of emphysema, cardiomegaly and hiatus hernia [34].

Few studies have been performed in the pediatric population. The first study that was published examined AI to find any abnormality in pediatric chest radiographs in a very-high-incidence population suspected of tuberculosis (113/119 images) [38]. The system reached reasonable performance with an AUC of 0.78 for correctly identifying abnormal regions in the image [38].

A recent study reported results for CAD4Kids [39], which identified pneumonia in children younger than 5 years on chest radiographs. Sensitivity of the system was 76%, with a specificity of 80%. ROC performance was 0.85, which was significantly lower than that of a reference observer who reached an AUC of 0.98.

In the study by Tang et al. [40], the authors tested their algorithm for the detection of pneumonia on pediatric data and reached an AUC of 0.92, which was lower than the AUC of 0.98 on adult data. After retraining on pediatric data, the AUC improved to 0.98. And after fine-tuning, their algorithm improved further and reached an AUC of 0.99 for classifying normal images versus images with pneumonia [40].

Chen et al. [41] developed an algorithm for detecting common abnormalities in chest radiographs specifically for children ages 1–17 years. The reference standard in the study was set by a pediatric pulmonologist and radiologist in consensus. The algorithm reached an accuracy of 72% for detecting bronchopneumonia and 85% for detecting lobar pneumonia on a test set of 531 images [41]. The authors found greater disagreement in children younger than 5 years. This might be explained by a greater variance of chest shape in younger children.

Zucker et al. [42] explored the performance of an AI algorithm to determine the Brasfield score in children with cystic fibrosis. Their model, trained on roughly 1,800 children and tested on 200 children, achieved comparable performance with five participating radiologists for the most sub-features in scoring cystic fibrosis components in chest radiographs according to the Brasfield score [42].

Currently, seven certified AI products can detect pneumonia and five products can detect pulmonary tuberculosis. CAD4TB, a commercial AI system for the automated detection of tuberculosis, is the only certified product that can be used in children age 4 years and older [17].

Several studies have been performed for automatic pneumothorax detection in chest radiography, of which few have been developed into commercial software. Hwang et al. [6] reported excellent performance of their algorithm, with a median AUC of 0.97 (range 0.92–0.99), for the detection of pulmonary malignancies, tuberculosis, pneumonia and pneumothorax. Most radiologists improved in performance after using the algorithm.

Park et al. [43] tested a convolutional neural network to detect pneumothorax after transthoracic needle biopsy. Their network was trained on 1,596 chest radiographs and reached an AUC of 0.90 for the detecting pneumothorax 3 h after needle biopsy. As a reference standard, they used a consensus read of two experienced radiologists [43].

The study of Rajpurkar et al. [34] also evaluated pneumothorax detection. In this study their algorithm reached an AUC of 0.944, compared to an average AUC of 0.940 of radiologists. No statistical difference was found between the algorithm and the radiologists. Unfortunately, the algorithm was not tested on an external dataset.

Chen et al. [41], who developed a deep learning system for the most common abnormalities in pediatric chest radiographs, also tested their algorithm for the detection of pneumothorax and reached an accuracy 86%.

Currently, 11 commercially available packages offer pneumothorax detection. According to the reported literature, none of the certified products has been trained on pediatric data. Therefore, performance level of these AI products might be lower in the pediatric population.

Another possible worthwhile application is the detection of lines and tubes in chest radiographs. Malposition of lines disrupts proper treatment, and clear visualization of the line can help to avoid repeat radiographs and reduce patient radiation dose.

Several studies have been performed for the detection of lines and tubes in chest radiographs [44‐49]. Generally the algorithms are quite good in classifying the presence versus absence of an endotracheal tube [50], but the systems perform worse when the exact position of the tip of the tube is sought.

Few studies for AI in the detection of lines and tubes have been performed in the pediatric population. In the study performed by Kao et al. [51], the authors developed an algorithm for detecting endotracheal tubes in neonatal chest radiographs. The authors evaluated the algorithm on 528 images with endotracheal tubes and 816 images without endotracheal tubes. The discriminant performance for detecting the existence of a tube reached an AUC of 0.94 [51]. The distance error of the tip of the tube was on average 1.89±2.01 mm [51].

In a study by Yi et al. [52], the authors used data with synthetic nasogastric tubes, endotracheal tubes and umbilical catheters to test their algorithm. The precision (i.e. true positives / [true positives + false positives]) for their algorithm was 0.80 [52]. According to the authors, their work can contribute to the development of a system that detects all lines and catheters in X-ray images and could be used to prioritize images that show malpositioned lines and request urgent review by the radiologist [52].

No commercial AI products are available that locate lines and tubes in chest radiographs. However, a single product on the market aims to optimize contrast for certain identification of lines and catheters (ClearRead Confirm; Riverain Technologies, Miamisburg, OH).

In CT, automatic lung nodule detection is the most widely studied application. The introduction of lung cancer screening has accelerated the quality and number of products for lung nodule detection in CT. Several studies have shown outstanding performance for the detection of lung nodules in CT [26, 28, 53, 54]; however, characterization of the nodules is a bigger issue. Many of the found nodules are small and benign. This has induced several approaches to characterize nodules and automatically estimate each nodule’s risk of malignancy [55, 56]. A recent study showed significantly improved performance of a convolutional neural network model for risk assessment of small nodules in CT compared with the Brock model, which is currently used for malignancy risk assessment, with an AUC of 0.90 compared with 0.87, respectively [55].

Eleven AI products are available for detecting pulmonary nodules in chest CT with the goal of finding early lung cancer. In the pediatric population, assessing pulmonary nodules in chest CT is not a clinical issue, and none of the products has been developed for the pediatric population. Nonetheless, these AI products might be useful for detecting pulmonary metastases, which is applicable in a specific pediatric population; none of products that is available, however, is specifically designed to detect pulmonary metastases.

Four products on the market focus on obstructive lung disease. These include chronic obstructive lung disease, asthma and cystic fibrosis. The latter two, especially, are applicable to children. Although CT examinations in asthma are not commonly performed, CT in cystic fibrosis patients is used to monitor the disease. AI software can detect minor attenuation differences in the lung, and thereby detect areas of air-trapping. Further, bronchopathic changes such as bronchial wall thickening and bronchiectasis can be detected and quantified. DeBoer et al. [57] showed that the number of airway counts on inspiratory high-resolution CT and the percentage of low-attenuation density on expiratory CT were significantly higher in children with cystic fibrosis.

A few products that can assess airway changes and air-trapping are available and should be usable in people with cystic fibrosis. However, no studies have been performed to assess the value of these products over visual assessment. Recently, Thirona B.V. (Nijmegen, The Netherlands) [58] obtained a patent for quantitative CT analysis of cystic fibrosis in children. They aim to develop an algorithm to identify early changes on CT scans in children with cystic fibrosis.

Several studies have examined the use of artificial intelligence in chest CT for the detection of pneumothorax. One example is the quick identification of pneumothoraces on chest CT [59]. This application reached a sensitivity of 100% and a specificity of 82.5%. The authors suggested that the tool could be used for triaging and to notify the radiologist about the detected urgent findings [59]. Complicating factors might be the presence of emphysema or bullae that could erroneously be seen as a pneumothorax. Other researchers tried to automatically quantify the volume of the pneumothorax. Rohrich et al. [60] showed excellent performance of their algorithm, with a Pearson correlation coefficient of 0.996 between the automated quantification method and the manual measurement. For the pediatric population, one study investigated a computer-aided volumetry scheme for quantifying pneumothoraces in chest CT. The study performed by Cai et al. [61] also showed excellent performance, with Pearson correlation coefficient of 0.99 between the manual and automated measurements. This is important because a larger-volume pneumothorax might trigger different treatment. At this point, no commercial AI product is available for detecting and quantifying pneumothoraces in chest CT.