Deep transfer learning for detection of placenta-mediated diseases from ultrasound images

Placenta-mediated diseases, such as preeclampsia (PE) and small-for-gestational-age (SGA) neonates, are associated with structural and functional changes in the placenta. While ultrasound is routinely performed during pregnancy, its potential for detecting such diseases using deep learning remains underexplored. This study presents a deep transfer learning approach using Convolutional Neural Networks (CNNs) to detect placenta-mediated diseases from ultrasound B-mode images of the human placenta ex vivo. This study is a secondary analysis of data from a prospective case-control study, where placentas were collected from women who delivered at BC Women's Hospital, Vancouver, Canada, from April 2018 to April 2020. Ultrasound data obtained from ex vivo human placentas were used to perform transfer learning using three deep learning architectures: DenseNet-121, ResNet-50, and InceptionV3, each pretrained on one of two datasets: ImageNet or RadImageNet, to detect SGA and/or PE outcome. The dataset consisted of 46 placentas obtained from 40 participants, of which 25 placentas were complicated by SGA and/or PE (including 12 SGA-only placentas, 2 PE placentas, and 11 placentas affected by both SGA and PE). Experimental results demonstrate that DenseNet-121, when pretrained with ImageNet, achieved the most promising performance, with accuracy of 0.76 ± 0.12, F1-score of 0.78 ± 0.09, and area under the receiver operating characteristic curve (AUC) of 0.86 ± 0.10. Among the RadImageNet pretrained models, ResNet-50 yielded an accuracy 0.75 ± 0.10, F1-score 0.77 ± 0.07, and AUC of 0.84 ± 0.09. Our proof-of-concept results demonstrate that standard B-mode ultrasound images of the placenta ex vivo contain sufficient information to distinguish between normal pregnancies and pregnancies affected by SGA and PE.