407 - Image Processing, Texture Features, and Machine Learning to Diagnose BPD in Preterm Infants

Publication Number: 1390.407

Karna Murthy, Ann & Robert H. Lurie Children's Hospital of Chicago, Chicago, IL, United States; Tomas Krepostman, Ann & Robert H. Lurie Children's Hospital of Chicago, Chicago, IL, United States; Yuan Si, Northwestern University, Evanston, IL, United States; Daizong Tian, Northwestern University, Chicago, IL, United States; Thrasyvoulos N. Pappas, Northwestern University, Evanston, IL, United States

Background: Bronchopulmonary dysplasia (BPD) is defined using gestational age and respiratory supports provided through 36 weeks' post-menstrual age (PMA). However, the predictive utility of the near-universally obtained chest radiograph (CXR) image to diagnose and predict infants' future risks of BPD is uncertain.

Objective: To determine whether objective descriptions of CXRs predict BPD accurately

Design/Methods: Non-anomalous infants < 32 weeks' gestation were included. The primary outcome was BPD, defined dichotomously if infants had grades 1-3 BPD using the 2019 Jensen definition. The main exposure was derived from infants' CXRs obtained between ages 5-14 d. After normalizing the CXR image intensity to the hepatic shadows, pixelated images were decomposed into 10 frequency subbands that include 4 spatial frequencies and 4 orientations using steerable filters (Figure 1). Texture features derived from subband image statistics were used to quantify CXR intensity and heterogeneity at different frequency scales. For image intensity, we relied on the low frequency subband. For heterogeneity, we relied on the two-horizontal frequency subbands to capture middle and high pulmonary frequency noise while filtering out contributions from radio-opaque bones and medical equipment. The diagnostic accuracy for grade 1-3 BPD is reported after applying 5-fold cross validation using linear support vector machine learning techniques after adjusting for gestational age and infant sex. The area under the receiver operating characteristic curve (AUC) is reported.

Results: Patients and their CXRs were (n=130) between from eligible infants were utilized. Of these, 50 (38.4%) had grade 1-3 BPD. The median gestational age and birth weight were 28 weeks' and 1080 g, respectively. Figures 2 and 3 show CXRs of representative infants stratified by BPD. Note that the energy in the mid- and high-frequency horizontal bands is increased for the infant with BPD, reflecting that the original CXR has increased heterogeneity. Lung heterogeneity, intensity and contrast measures from CXRs accurately predict BPD at 36 weeks' PMA adjusted for gestational age and sex (sensitivity=0.88; specificity=0.80; negative predictive value=0.82; positive predictive value=0.88; AUC= 0.91). Thus, the overall accuracy using these techniques was 84%.

Conclusion(s): Image processing of CXRs show promise to diagnose and predict BPD objectively in infants born < 32 weeks' gestation. It remains uncertain and an ongoing direction whether these techniques, in conjunction with patient characteristics, can predict respiratory outcomes and/or severity of BPD.

Range of Frequencies for CXR Decomposition into Subbands
This figure describes the decomposition of CXR images into subbands of different non-overlapping spatial frequencies. The sum of all the subbands would reconstruct the original image/CXR. Our analysis focused on the horizontal subbands which filter out the radio-opaque bones.

Representative Infant WITHOUT BPD, their CXR, and Horizontal Frequency Subbands
The left panel is the original CXR; the middle panel is the mid-frequency subband; the right panel is the high frequency subband. Both the middle and right panels are deconstructed from the original digitized radiograph.

Representative Infant WITH BPD, their CXR, and Horizontal Frequency Subbands
The left panel is the original CXR; the middle panel is the mid-frequency subband; the right panel is the high frequency subband. Both the middle and right panels are deconstructed from the original digitized radiograph. This picture depicts displays more intensity relative to the infant without BPD