Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PRISM Lite: A lightweight model for interactive 3D placenta segmentation in ultrasound

Hao Li, Baris Oguz, Gabriel Arenas, Xing Yao, Jiacheng Wang, Alison Pouch, Brett Byram, Nadav Schwartz, Ipek Oguz

TL;DR

Placental volume from 3D ultrasound is a valuable metric for fetal health but automated placenta segmentation can be unreliable in challenging images. This work introduces PRISM Lite, a lightweight interactive segmentation framework that starts from an automated placenta mask and uses a two-path network with sparse prompts and iterative learning over 11 steps to refine segmentation. PRISM Lite achieves near state-of-the-art accuracy (Dice ~0.98, NSD ~0.999) with about 0.13M parameters and ~0.75 s CPU inference, vastly reducing computational demands compared to larger models like PRISM. The results show strong performance and robustness to imperfect initial masks, highlighting potential for real-time deployment on mobile devices or ultrasound consoles in low-resource settings.

Abstract

Placenta volume measured from 3D ultrasound (3DUS) images is an important tool for tracking the growth trajectory and is associated with pregnancy outcomes. Manual segmentation is the gold standard, but it is time-consuming and subjective. Although fully automated deep learning algorithms perform well, they do not always yield high-quality results for each case. Interactive segmentation models could address this issue. However, there is limited work on interactive segmentation models for the placenta. Despite their segmentation accuracy, these methods may not be feasible for clinical use as they require relatively large computational power which may be especially prohibitive in low-resource environments, or on mobile devices. In this paper, we propose a lightweight interactive segmentation model aiming for clinical use to interactively segment the placenta from 3DUS images in real-time. The proposed model adopts the segmentation from our fully automated model for initialization and is designed in a human-in-the-loop manner to achieve iterative improvements. The Dice score and normalized surface Dice are used as evaluation metrics. The results show that our model can achieve superior performance in segmentation compared to state-of-the-art models while using significantly fewer parameters. Additionally, the proposed model is much faster for inference and robust to poor initial masks. The code is available at https://github.com/MedICL-VU/PRISM-placenta.

PRISM Lite: A lightweight model for interactive 3D placenta segmentation in ultrasound

TL;DR

Placental volume from 3D ultrasound is a valuable metric for fetal health but automated placenta segmentation can be unreliable in challenging images. This work introduces PRISM Lite, a lightweight interactive segmentation framework that starts from an automated placenta mask and uses a two-path network with sparse prompts and iterative learning over 11 steps to refine segmentation. PRISM Lite achieves near state-of-the-art accuracy (Dice ~0.98, NSD ~0.999) with about 0.13M parameters and ~0.75 s CPU inference, vastly reducing computational demands compared to larger models like PRISM. The results show strong performance and robustness to imperfect initial masks, highlighting potential for real-time deployment on mobile devices or ultrasound consoles in low-resource settings.

Abstract

Placenta volume measured from 3D ultrasound (3DUS) images is an important tool for tracking the growth trajectory and is associated with pregnancy outcomes. Manual segmentation is the gold standard, but it is time-consuming and subjective. Although fully automated deep learning algorithms perform well, they do not always yield high-quality results for each case. Interactive segmentation models could address this issue. However, there is limited work on interactive segmentation models for the placenta. Despite their segmentation accuracy, these methods may not be feasible for clinical use as they require relatively large computational power which may be especially prohibitive in low-resource environments, or on mobile devices. In this paper, we propose a lightweight interactive segmentation model aiming for clinical use to interactively segment the placenta from 3DUS images in real-time. The proposed model adopts the segmentation from our fully automated model for initialization and is designed in a human-in-the-loop manner to achieve iterative improvements. The Dice score and normalized surface Dice are used as evaluation metrics. The results show that our model can achieve superior performance in segmentation compared to state-of-the-art models while using significantly fewer parameters. Additionally, the proposed model is much faster for inference and robust to poor initial masks. The code is available at https://github.com/MedICL-VU/PRISM-placenta.
Paper Structure (4 sections, 4 figures, 2 tables)

This paper contains 4 sections, 4 figures, 2 tables.

Table of Contents

  1. INTRODUCTION
  2. Materials and Methods
  3. Results
  4. Conclusion

Figures (4)

  • Figure 1: Top, the proposed 3D interactive segmentation framework, illustrated in 2D. Bottom, warped scribble generation. Briefly, they are generated by breaking non-warped scribbles into line segments and warping li2024inter.
  • Figure 2: Network architecture of the proposed lightweight model. The numbers of output channels are marked.
  • Figure 3: Qualitative results. Orange arrows highlight segmentation errors. PRISM Lite rapidly corrects these errors.
  • Figure 4: (a) Performance using the automated model result as initial mask for PRISM Lite. (b-d) Performance using polluted initial masks. (e) illustrates these polluted masks in white, overlaid with the raw automated model mask in red.