Li, H.; Oguz, B.; Arenas, G.; Yao, X.; Wang, J.; Pouch, A.; Byram, B.; Schwartz, N.; Oguz, I. “Interactive Segmentation Model for Placenta Segmentation from 3D Ultrasound Images.” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Volume 15186 LNCS, 2025, pp. 132-142, DOI: 10.1007/978-3-031-73647-6_13.
Measuring placenta volume from 3D ultrasound images is important for predicting pregnancy outcomes, but manually outlining the placenta in these images is time-consuming and costly. While automated methods can segment the placenta, they often aren’t reliable enough for consistent use. Recently, interactive deep learning models, inspired by tools like the Segment Anything Model (SAM), have been applied to medical images. These models allow users to provide prompts, helping the model identify and segment the target area, which could make them useful in practice.
However, existing interactive models are not specifically designed for the unique challenges of 3D ultrasound images, which are often noisy. This study tested the performance of several state-of-the-art 3D interactive segmentation models against a “human-in-the-loop” approach, where a person helps guide the model during segmentation. The evaluation used several metrics, including the Dice score, which measures how closely the model’s output matches the manual annotation. A Dice score of 0.95 was considered a successful result.
The findings show that the human-in-the-loop model achieved this high level of accuracy and performed well even with a limited number of user prompts, making it both effective and efficient for segmenting the placenta. The code for this study is available online for further use and testing at https://github.com/MedICL-VU/PRISM-placenta.
Fig (1)
(a) The 3D interactive segmentation model (PRISM [10]), illustrated in 2D. (b) Prompt sampling for positive (left) and negative prompts (right). To mimic human behavior, we sample prompts from the FN and FP regions of the current segmentation at each iteration. The initial sampling only has positive prompts.