Sequential Spatial-Temporal Network for Interpretable Automatic Ultrasonic Assessment of Fetal Head during labor

TL;DR

The paper tackles automatic, interpretable assessment of fetal head descent during labor using intrapartum ultrasound by focusing on the ISUOG metrics $AoP$ and $HSD$. It introduces the Sequential Spatial-Temporal Network (SSTN), a first interpretable model for intrapartum ultrasound video, which sequentially identifies planes, segments anatomical structures, and detects landmarks to compute $AoP$ and $HSD$ while leveraging temporal context. Through multitask supervision and a three-stage architecture (feature enhancement, Video Swin Transformer encoder, and a ResConv-UpConv decoder), SSTN achieves state-of-the-art performance, reducing $\Delta$AoP by 18% and $\Delta$HSD by 22% compared to baselines. The approach demonstrates improved robustness and interpretability, with potential for clinical deployment in labor assessments and guidance for future research in ultrasound video analysis.

Abstract

The intrapartum ultrasound guideline established by ISUOG highlights the Angle of Progression (AoP) and Head Symphysis Distance (HSD) as pivotal metrics for assessing fetal head descent and predicting delivery outcomes. Accurate measurement of the AoP and HSD requires a structured process. This begins with identifying standardized ultrasound planes, followed by the detection of specific anatomical landmarks within the regions of the pubic symphysis and fetal head that correlate with the delivery parameters AoP and HSD. Finally, these measurements are derived based on the identified anatomical landmarks. Addressing the clinical demands and standard operation process outlined in the ISUOG guideline, we introduce the Sequential Spatial-Temporal Network (SSTN), the first interpretable model specifically designed for the video of intrapartum ultrasound analysis. The SSTN operates by first identifying ultrasound planes, then segmenting anatomical structures such as the pubic symphysis and fetal head, and finally detecting key landmarks for precise measurement of HSD and AoP. Furthermore, the cohesive framework leverages task-related information to improve accuracy and reliability. Experimental evaluations on clinical datasets demonstrate that SSTN significantly surpasses existing models, reducing the mean absolute error by 18% for AoP and 22% for HSD.

Figures (4)

• Figure 1: The illustration of fetal head assessment by AoP and HSD. Free-head positioning to measure the AoP (a) and HSD (b). The ultrasound biometric (c) is measured by following the protocols in (a) and (b).
• Figure 2: The proposed SSTN includes a feature enhancement block, a Video Swin Transformer as the encoder, and a decoder with ResConvs and UpConvs for optimized feature extraction. The Sequential Multi-task Learning framework uses a weighted loss function to balance classification, segmentation, and landmark detection tasks.
• Figure 3: The features extracted by various models during the classification task can be visualized using Gradient-weighted Class Activation Mapping (Grad-CAM).
• Figure 4: Segmentation (left) and keypoint (right) prediction results for various methods. In segmentation, the blue region marks the pubic symphysis, yellow indicates the fetal head, and the red curve represents the ground truth. For keypoint predictions, red points denote the ground truth, while green points show predicted keypoints.