Automated detection of acute respiratory distress syndrome from chest X-Rays using Directionality Measure and deep learning features

Abstract

Acute respiratory distress syndrome (ARDS) is a life-threatening lung injury with global prevalence and high mortality. Chest x-rays (CXR) are critical in the early diagnosis and treatment of ARDS. However, imaging findings may not result in proper identification of ARDS due to a number of reasons, including nonspecific appearance of radiological features, ambiguity in a patient's case due to the pathological stage of the disease, and poor inter-rater reliability from interpretations of CXRs by multiple clinical experts. This study demonstrates the potential capability of methodologies in artificial intelligence, machine learning, and image processing to overcome these challenges and quantitatively assess CXRs for presence of ARDS. We propose and describe Directionality Measure, a novel feature engineering technique used to capture the "cloud-like" appearance of diffuse alveolar damage as a mathematical concept. This study also examines the effectiveness of using an off-the-shelf, pretrained deep learning model as a feature extractor in addition to standard features extracted from the histogram and gray-level co-occurrence matrix (GLCM). Data was collected from hospitalized patients at Michigan Medicine's intensive care unit and the cohort's inclusion criteria was specifically designed to be representative of patients at risk of developing ARDS. Multiple machine learning models were used to evaluate these features with 5-fold cross-validation and the final performance was reported on a hold-out, temporally distinct test set. With AdaBoost, Directionality Measure achieved an accuracy of 78% and AUC of 74% - outperforming classification results using features from the histogram (75% accuracy and 73% AUC), GLCM (76% accuracy and 73% AUC), and ResNet-50 (77% accuracy and 73% AUC). Further experimental results demonstrated that using all feature sets in combination achieved the best overall performance, yielding an accuracy of 83% and AUC of 79% with AdaBoost. These results demonstrate the potential capability of using the proposed methodologies to complement current clinical analysis for detection of ARDS from CXRs.

Keywords: Acute respiratory distress syndrome; Chest X-ray images; Deep learning; Image processing; Machine learning.

Fig. 1.

Examples of chest x-rays from the Michigan Medicine dataset that were annotated as consistent with ARDS or inconsistent with ARDS based on the reviews of multiple clinical experts. Chest x-rays (a), (b), (c) demonstrate the findings of ARDS, which are bilateral airspace disease not explained by effusions, lobar/lung collapse, or nodules. These findings may (b) or may not (a, c) be uniform across both lung fields. Chest x-rays (d), (e), (f) do not demonstrate clear findings of ARDS, either because the lung fields lack clear airspace disease (d) or the disease that is present is unilateral (e, f).

Fig. 2.

Label generation of chest x-ray scans. (a) Multiple expert clinicians were asked to independently review patients’ CXR and determine if any individuals had ARDS. Clinicians were also rated their confidence of the diagnosis as equivocal, slightly confident, moderately confident, or highly confident. (b) The diagnosis and confidence were converted to a scale between 1-8. The final label was generated from aggregating these reviews to ensure correctness and consistency of the diagnosis. A label of −1 (non-ARDS) was assigned if the averaged review was below or equal to 4.5, and a label of 1 (ARDS) was assigned if the averaged review was above 4.5

Fig. 3.

Directionality Measure applied to CXRs (a) from a patient diagnosed with ARDS and (b) from a non-ARDS diagnosis. The /original CXR is shown in the upper left, G in the lower left, H in the upper right, and G · H in the lower right for both figures.