Pilot study of machine learning for detection of placenta accreta spectrum

Abstract

Antenatal suspicion of placenta accreta spectrum (PAS) is a prerequisite for timely referral to a multidisciplinary team. While ultrasound is the mainstay for antenatal suspicion of PAS, it is operator dependent. Furthermore, a patient's a-priori clinical risk for PAS influences the interpretation of PAS-targeted studies, which is a limitation when designing research studies to objectively assess diagnostic techniques1. Radiomics, an analysis technique that extracts quantifiable information from images, transforms medical imaging into data that can be used for deep learning. The objective of this pilot study was to utilize machine learning to detect the basic anatomical features assessed during diagnosis of PAS, including the placenta, uterus and bladder. We conducted a retrospective review of ultrasound studies conducted at 26–32 weeks' gestation in subjects with placenta previa on ultrasound and PAS on final uterine pathology at a single referral center between 2020–20222. For each subject, we selected grayscale images obtained in the sagittal plane, including the placenta, bladder and lower uterine segment. As this was a pilot study, we excluded axial and color Doppler images, as well as images without a full bladder. The images used were obtained as part of routine clinical care. After images were selected, we annotated the placenta, bladder and uterus using LabelMe software (Massachusetts Institute of Technology, Cambridge, MA, USA). The images were then divided randomly into training, validation and testing sets. Once the images were annotated and divided into sets, we deployed three algorithms (DeepLabV33, UNet4 and Inception-UNet5) to generate machine segmentation of the same anatomical features (bladder, uterus and placenta). We further integrated semi-supervized learning algorithms using an exponential moving average-mean teacher (EMA-MT) model6. Full methodology is given in Appendix S1, and our algorithms are open source for further research. Dice's coefficient was the primary outcome, indicating the strength of agreement between the human-annotated and machine-annotated segmentation images (scored 0.0–1.0, where 0.0 represents no overlap and 1.0 is perfect agreement)7. There were 37 subjects included in the study (Table 1). Notably, nine (24.3%) had International Federation of Gynecology and Obstetrics8 (FIGO) Grade-I PAS, 16 (43.2%) had Grade-II PAS and 12 (32.4%) had Grade-III PAS. From these 37 subjects, 357 sonographic frames met the inclusion criteria and were annotated (including 88 (24.6%) Grade I, 115 (32.2%) Grade II, 141 (39.5%) Grade III, and 13 (3.6%) of unknown grade that were not captured during the image anonymization process). Variants of the DeepLabV3 model best segmented the placenta, with the best model (DeepLab-CE-EMA-MT) obtaining a Dice coefficient (± SD) of 0.86 ± 0.12 (Table 2). The maximum Dice coefficients (± SD) were lower and had larger fluctuations for the bladder (0.77 ± 0.39) and uterus (0.61 ± 0.30). The Dice coefficients for placental segmentation observed in individuals with PAS included in this study are similar to coefficients published previously for normal placentas (0.87 ± 0.10 for anterior placenta; 0.80 ± 0.13 for posterior placenta)9. The next steps in research are to assess whether machine learning modeling can differentiate between normal and pathologically adherent placentas, and to prospectively validate these results with lower grade placentation. A surprising finding was the low segmentation accuracy for the bladder. Bladder segmentation accuracy has not been reported previously, likely in part because it is not a component of routine assessment in obstetric ultrasound. A challenge to machine learning segmentation of the bladder includes the presence of amniotic fluid, which has similar echolucency to the bladder. In reviewing the testing slides in our study, the reason for low segmentation accuracy was clear: amniotic fluid was modeled as part of bladder segmentation. Future models will need to distinguish the two in order to improve bladder segmentation. In our retrospective study of patients with histological PAS, we demonstrated that machine learning models may readily segment the placenta, with lower segmentation accuracy for the bladder and uterus. This study was funded by the Artificial Intelligence Spark award, sponsored by Duke AI Health and Duke Center for Artificial Intelligence in Radiology, an intramural award to bridge clinical medicine and artificial intelligence. The funding source had no role in the study design. Algorithms used are made opensource and available for further research via the PyTorch library: https://pytorch.org. Appendix S1 Full methodology Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.