Interactive Segmentation Model for Placenta Segmentation from 3D Ultrasound images

TL;DR

This work addresses the challenge of robust placenta segmentation from 3D ultrasound by evaluating public 3D interactive segmentation models against a human-in-the-loop baseline. It demonstrates that PRISM, when guided by 2D scribble prompts and iterative refinements, can achieve near-human Dice scores (approximately 0.97) with NSD near 0.999, outperforming other interactive methods that rely on single prompts. The study also shows that performance improves quickly in early iterations, is resilient to sparse prompting, and remains effective with smaller training sets, indicating strong practical potential for annotation support and clinical workflows. Overall, PRISM offers an effective and efficient solution for interactive placenta segmentation in 3DUS, with code availability and avenues for future efficiency enhancements through pretraining.

Abstract

Placenta volume measurement from 3D ultrasound images is critical for predicting pregnancy outcomes, and manual annotation is the gold standard. However, such manual annotation is expensive and time-consuming. Automated segmentation algorithms can often successfully segment the placenta, but these methods may not consistently produce robust segmentations suitable for practical use. Recently, inspired by the Segment Anything Model (SAM), deep learning-based interactive segmentation models have been widely applied in the medical imaging domain. These models produce a segmentation from visual prompts provided to indicate the target region, which may offer a feasible solution for practical use. However, none of these models are specifically designed for interactively segmenting 3D ultrasound images, which remain challenging due to the inherent noise of this modality. In this paper, we evaluate publicly available state-of-the-art 3D interactive segmentation models in contrast to a human-in-the-loop approach for the placenta segmentation task. The Dice score, normalized surface Dice, averaged symmetric surface distance, and 95-percent Hausdorff distance are used as evaluation metrics. We consider a Dice score of 0.95 a successful segmentation. Our results indicate that the human-in-the-loop segmentation model reaches this standard. Moreover, we assess the efficiency of the human-in-the-loop model as a function of the amount of prompts. Our results demonstrate that the human-in-the-loop model is both effective and efficient for interactive placenta segmentation. The code is available at \url{https://github.com/MedICL-VU/PRISM-placenta}.

Figures (5)

• Figure 1: (a) The 3D interactive segmentation model (PRISM li2024prism), 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.
• Figure 2: Top: Centerline scribble generated from a longitudinal slice. Bottom: Boundary scribble generated from a transverse view. The binary mask can be derived from either FP or FN masks to generate negative and positive scribbles, respectively.
• Figure 3: Qualitative results of compared methods for placenta segmentation.
• Figure 4: PRISM segmentation performance with different scribbles across iterations. (a) Dice, (b) NSD, (c) ASD, (d) HD95. The mean values (lines) and 95% confidence intervals (shades) are shown. T and L denote transverse and longitudinal views, respectively.
• Figure 5: (a) Dice per iteration for sparse sampling. (b) Prompt-to-volume ratio at each iteration for sparse sampling. The prompt-to-volume ratio is the ratio of the total number of scribble voxels to the number of placenta voxels in the ground truth. (c) Dice score on different training set size, without changing the test set.