Deep learning for classification of pediatric chest radiographs by WHO's standardized methodology

A comprehensive segmentation of chest X-ray improves deep learning-based WHO radiologically confirmed pneumonia diagnosis in children

OBJECTIVES

To investigate a comprehensive segmentation of chest X-ray (CXR) in promoting deep learning-based World Health Organization's (WHO) radiologically confirmed pneumonia diagnosis in children.

METHODS

A total of 4400 participants between January 2016 and June 2021were identified for a cross-sectional study and divided into primary endpoint pneumonia (PEP), other infiltrates, and normal groups according to WHO's diagnostic criteria. The CXR was divided into six segments of left lung, right lung, mediastinum, diaphragm, ext-left lung, and ext-right lung by adopting the RA-UNet. To demonstrate the benefits of lung field segmentation in pneumonia diagnosis, the segmented images and images that were not segmented, which constituted seven segmentation combinations, were fed into the CBAM-ResNet under a three-category classification comparison. The interpretability of the CBAM-ResNet for pneumonia diagnosis was also performed by adopting a Grad-CAM module.

RESULTS

The RA-UNet achieved a high spatial overlap between manual and automatic segmentation (averaged DSC = 0.9639). The CBAM-ResNet when fed with the six segments achieved superior three-category diagnosis performance (accuracy = 0.8243) over other segmentation combinations and deep learning models under comparison, which was increased by around 6% in accuracy, precision, specificity, sensitivity, F1-score, and around 3% in AUC. The Grad-CAM could capture the pneumonia lesions more accurately, generating a more interpretable visualization and enhancing the superiority and reliability of our study in assisting pediatric pneumonia diagnosis.

CONCLUSIONS

The comprehensive segmentation of CXR could improve deep learning-based pneumonia diagnosis in childhood with a more reasonable WHO's radiological standardized pneumonia classification instead of conventional dichotomous bacterial pneumonia and viral pneumonia.

CLINICAL RELEVANCE STATEMENT

The comprehensive segmentation of chest X-ray improves deep learning-based WHO confirmed pneumonia diagnosis in children, laying a strong foundation for the potential inclusion of computer-aided pediatric CXR readings in precise classification of pneumonia and PCV vaccine trials efficacy in children.

KEY POINTS

• The chest X-ray was comprehensively segmented into six anatomical structures of left lung, right lung, mediastinum, diaphragm, ext-left lung, and ext-right lung. • The comprehensive segmentation improved the three-category classification of primary endpoint pneumonia, other infiltrates, and normal with an increase by around 6% in accuracy, precision, specificity, sensitivity, F1-score, and around 3% in AUC. • The comprehensive segmentation gave rise to a more accurate and interpretable visualization results in capturing the pneumonia lesions.

A machine learning approach to explore individual risk factors for tuberculosis treatment non-adherence in Mukono district

Despite the availability and implementation of well-known efficacious interventions for tuberculosis treatment by the Ministry of Health, Uganda (MoH), treatment non-adherence persists. Moreover, identifying a specific tuberculosis patient at risk of treatment non-adherence is still a challenge. Thus, this retrospective study, based on a record review of 838 tuberculosis patients enrolled in six health facilities, presents, and discusses a machine learning approach to explore the individual risk factors predictive of tuberculosis treatment non-adherence in the Mukono district, Uganda. Five classification machine learning algorithms, logistic regression (LR), artificial neural networks (ANN), support vector machines (SVM), random forest (RF), and AdaBoost were trained, and evaluated by computing their accuracy, F1 score, precision, recall, and the area under the receiver operating curve (AUC) through the aid of a confusion matrix. Of the five developed and evaluated algorithms, SVM (91.28%) had the highest accuracy (AdaBoost, 91.05% performed better than SVM when AUC is considered as evaluation parameter). Looking at all five evaluation parameters globally, AdaBoost is quite on par with SVM. Individual risk factors predictive of non-adherence included tuberculosis type, GeneXpert results, sub-country, antiretroviral status, contacts below 5 years, health facility ownership, sputum test results at 2 months, treatment supporter, cotrimoxazole preventive therapy (CPT) dapsone status, risk group, patient age, gender, middle and upper arm circumference, referral, positive sputum test at 5 and 6 months. Therefore, machine learning techniques, specifically classification types, can identify patient factors predictive of treatment non-adherence and accurately differentiate between adherent and non-adherent patients. Thus, tuberculosis program management should consider adopting the classification machine learning techniques evaluated in this study as a screening tool for identifying and targeting suited interventions to these patients.

Optimising computer aided detection to identify intra-thoracic tuberculosis on chest x-ray in South African children

Diagnostic tools for paediatric tuberculosis remain limited, with heavy reliance on clinical algorithms which include chest x-ray. Computer aided detection (CAD) for tuberculosis on chest x-ray has shown promise in adults. We aimed to measure and optimise the performance of an adult CAD system, CAD4TB, to identify tuberculosis on chest x-rays from children with presumptive tuberculosis. Chest x-rays from 620 children <13 years enrolled in a prospective observational diagnostic study in South Africa, were evaluated. All chest x-rays were read by a panel of expert readers who attributed each with a radiological reference of either 'tuberculosis' or 'not tuberculosis'. Of the 525 chest x-rays included in this analysis, 80 (40 with a reference of 'tuberculosis' and 40 with 'not tuberculosis') were allocated to an independent test set. The remainder made up the training set. The performance of CAD4TB to identify 'tuberculosis' versus 'not tuberculosis' on chest x-ray against the radiological reference read was calculated. The CAD4TB software was then fine-tuned using the paediatric training set. We compared the performance of the fine-tuned model to the original model. Our findings were that the area under the receiver operating characteristic curve (AUC) of the original CAD4TB model, prior to fine-tuning, was 0.58. After fine-tuning there was an improvement in the AUC to 0.72 (p = 0.0016). In this first-ever description of the use of CAD to identify tuberculosis on chest x-ray in children, we demonstrate a significant improvement in the performance of CAD4TB after fine-tuning with a set of well-characterised paediatric chest x-rays. CAD has the potential to be a useful additional diagnostic tool for paediatric tuberculosis. We recommend replicating the methods we describe using a larger chest x-ray dataset from a more diverse population and evaluating the potential role of CAD to replace a human-read chest x-ray within treatment-decision algorithms for paediatric tuberculosis.

The development of a novel natural language processing tool to identify pediatric chest radiograph reports with pneumonia

Objective

Chest radiographs are frequently used to diagnose community-acquired pneumonia (CAP) for children in the acute care setting. Natural language processing (NLP)-based tools may be incorporated into the electronic health record and combined with other clinical data to develop meaningful clinical decision support tools for this common pediatric infection. We sought to develop and internally validate NLP algorithms to identify pediatric chest radiograph (CXR) reports with pneumonia.

Materials and methods

We performed a retrospective study of encounters for patients from six pediatric hospitals over a 3-year period. We utilized six NLP techniques: word embedding, support vector machines, extreme gradient boosting (XGBoost), light gradient boosting machines Naïve Bayes and logistic regression. We evaluated their performance of each model from a validation sample of 1,350 chest radiographs developed as a stratified random sample of 35% admitted and 65% discharged patients when both using expert consensus and diagnosis codes.

Results

Of 172,662 encounters in the derivation sample, 15.6% had a discharge diagnosis of pneumonia in a primary or secondary position. The median patient age in the derivation sample was 3.7 years (interquartile range, 1.4-9.5 years). In the validation sample, 185/1350 (13.8%) and 205/1350 (15.3%) were classified as pneumonia by content experts and by diagnosis codes, respectively. Compared to content experts, Naïve Bayes had the highest sensitivity (93.5%) and XGBoost had the highest F1 score (72.4). Compared to a diagnosis code of pneumonia, the highest sensitivity was again with the Naïve Bayes (80.1%), and the highest F1 score was with the support vector machine (53.0%).

Conclusion

NLP algorithms can accurately identify pediatric pneumonia from radiography reports. Following external validation and implementation into the electronic health record, these algorithms can facilitate clinical decision support and inform large database research.

Artificial intelligence-based clinical decision support in pediatrics

Machine learning models may be integrated into clinical decision support (CDS) systems to identify children at risk of specific diagnoses or clinical deterioration to provide evidence-based recommendations. This use of artificial intelligence models in clinical decision support (AI-CDS) may have several advantages over traditional "rule-based" CDS models in pediatric care through increased model accuracy, with fewer false alerts and missed patients. AI-CDS tools must be appropriately developed, provide insight into the rationale behind decisions, be seamlessly integrated into care pathways, be intuitive to use, answer clinically relevant questions, respect the content expertise of the healthcare provider, and be scientifically sound. While numerous machine learning models have been reported in pediatric care, their integration into AI-CDS remains incompletely realized to date. Important challenges in the application of AI models in pediatric care include the relatively lower rates of clinically significant outcomes compared to adults, and the lack of sufficiently large datasets available necessary for the development of machine learning models. In this review article, we summarize key concepts related to AI-CDS, its current application to pediatric care, and its potential benefits and risks. IMPACT: The performance of clinical decision support may be enhanced by the utilization of machine learning-based algorithms to improve the predictive performance of underlying models. Artificial intelligence-based clinical decision support (AI-CDS) uses models that are experientially improved through training and are particularly well suited toward high-dimensional data. The application of AI-CDS toward pediatric care remains limited currently but represents an important area of future research.

Parental Perceptions on Use of Artificial Intelligence in Pediatric Acute Care

BACKGROUND

Family engagement is critical in the implementation of artificial intelligence (AI)-based clinical decision support tools, which will play an increasing role in health care in the future. We sought to understand parental perceptions of computer-assisted health care of children in the emergency department (ED).

METHODS

We conducted a population-weighted household panel survey of parents with minor children in their home in a large US city to evaluate perceptions of the use of computer programs for the care of children with respiratory illness. We identified demographics associated with discomfort with AI using survey-weighted logistic regression.

RESULTS

Surveys were completed by 1620 parents (panel response rate = 49.7%). Most respondents were comfortable with the use of computer programs to determine the need for antibiotics (77.6%) or bloodwork (76.5%), and to interpret radiographs (77.5%). In multivariable analysis, Black non-Hispanic parents reported greater discomfort with AI relative to White non-Hispanic parents (odds ratio [OR] 1.67, 95% confidence interval [CI] 1.03-2.70) as did younger parents (18-25 years) relative to parents ≥46 years (OR 2.48, 95% CI 1.31-4.67). The greatest perceived benefits of computer programs were finding something a human would miss (64.2%, 95% CI 60.9%-67.4%) and obtaining a more rapid diagnosis (59.6%; 56.2%-62.9%). Areas of greatest concern were diagnostic errors (63.0%, 95% CI 59.6%-66.4%), and recommending incorrect treatment (58.9%, 95% CI 55.5%-62.3%).

CONCLUSIONS

Parents were generally receptive to computer-assisted management of children with respiratory illnesses in the ED, though reservations emerged. Black non-Hispanic and younger parents were more likely to express discomfort about AI.

Pediatric chest radiograph interpretation: how far has artificial intelligence come? A systematic literature review

Most artificial intelligence (AI) studies have focused primarily on adult imaging, with less attention to the unique aspects of pediatric imaging. The objectives of this study were to (1) identify all publicly available pediatric datasets and determine their potential utility and limitations for pediatric AI studies and (2) systematically review the literature to assess the current state of AI in pediatric chest radiograph interpretation. We searched PubMed, Web of Science and Embase to retrieve all studies from 1990 to 2021 that assessed AI for pediatric chest radiograph interpretation and abstracted the datasets used to train and test AI algorithms, approaches and performance metrics. Of 29 publicly available chest radiograph datasets, 2 datasets included solely pediatric chest radiographs, and 7 datasets included pediatric and adult patients. We identified 55 articles that implemented an AI model to interpret pediatric chest radiographs or pediatric and adult chest radiographs. Classification of chest radiographs as pneumonia was the most common application of AI, evaluated in 65% of the studies. Although many studies report high diagnostic accuracy, most algorithms were not validated on external datasets. Most AI studies for pediatric chest radiograph interpretation have focused on a limited number of diseases, and progress is hindered by a lack of large-scale pediatric chest radiograph datasets.