Automatic segmentation of placenta for quantitative ultrasound analysis

Quantitative ultrasound (QUS) holds promise for non-invasive placental tissue characterization and disease detection, yet its clinical application is hindered by the effort required for placental segmentation to identify a region-of-interest (ROI) for calculation. Manual segmentation is time-consuming and prone to variability due to the irregular boundaries of the placental tissue. This study aims to develop an automatic placental identification and segmentation model to facilitate QUS integration into clinical practice, alleviate clinician workload, and enable real-time feedback on QUS quality. We employed Mask R-CNN within the Detectron2 framework to automate placental segmentation using a dataset of 149 B-mode ultrasound images annotated by a medical image specialist (M.D.) covering various trimesters. Inference on a validation subset (30 images) yielded an average Dice Similarity Coefficient (DSC) of 0.863. Our model gave quality predictions for a majority of the test images with 57% of segmentations achieving DSC ≥ 0.85 and 40% with DSC ≥ 0.90. The model’s low inference time of approximately 55 ms per iteration supports near real-time QUS processing. Future work will focus on expanding the dataset, optimizing the loss function, and addressing edge effects on QUS to further enhance segmentation quality.