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

summary

The gist

The Sequential Spatial-Temporal Network (SSTN) is introduced as the first interpretable model specifically designed for analyzing intrapartum ultrasound videos to automatically assess fetal head…

In short

The Sequential Spatial-Temporal Network (SSTN) is an interpretable model designed to automatically assess fetal head descent during labor from ultrasound videos. It achieves high accuracy by identifying ultrasound planes, segmenting structures like the fetal head and pubic symphysis, and detecting landmarks for precise measurement of Head Symphysis Distance (HSD) and Angle of Progression (AoP). The model significantly improves existing methods.

Key concepts

Head Symphysis Distance (HSD)
This is a key measurement derived from the ultrasound video that quantifies the distance between the fetal head and the pubic symphysis. Accurately measuring HSD is crucial for tracking fetal descent during labor, and this model focuses on precise detection of this specific distance.
Angle of Progression (AoP)
AoP measures the angle at which the fetal head is progressing relative to a reference line in the ultrasound image. This angle is an important indicator of labor progress, and the SSTN model aims to accurately detect this angle for reliable assessment.
Weighted Loss Function
The SSTN uses a combined loss function (Losstotal) that balances three different tasks: classification, segmentation, and landmark detection. This function assigns different weights (alpha, beta, gamma) to each task, ensuring the model learns effectively across all necessary components simultaneously.
Video Swin Transformer (VST)
The VST encoder is a component of the SSTN architecture that processes video sequences. It is designed to capture rich temporal and spatial information from consecutive frames in the ultrasound video, helping the model understand how fetal head descent changes over time.

Terminology used across episodes

This episode discusses

The paper

Sequential Spatial-Temporal Network for Interpretable Automatic Ultrasonic Assessment of Fetal Head during labor · Read on arXiv

Jie Gan, Zhuonan Liang, Jianan Fan, Lisa Mcguire, Caterina Watson, Jacqueline Spurway, Jillian Clarke, *Weidong Cai

School of Computer Science, The University of Sydney, NSW, Australia · Sydney School of Health Sciences, The University of Sydney, NSW, Australia · School of Science, Edith Cowan University, WA, Australia · Medical Imaging Department at Orange Hospital

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Sequential Spatial-Temporal Network for Interpretable Automatic Ultrasonic Assessment of Fetal Head during labor".

Dev: Here's the gist of the paper. The Sequential Spatial-Temporal Network (SSTN) is introduced as the first interpretable model specifically designed for analyzing intrapartum ultrasound videos to automatically assess fetal head descent during labor The SSTN significantly surpasses existing models, reducing the mean absolute error by 18% for AoP and 22% for HSD Objective of the Assessment The assessment relies on identifying standardized ultrasound planes, segmenting anatomical structures like the pubic symphysis and fetal head, and detecting key landmarks for precise measurement of Head Symphysis Distance (HSD) and Angle of Progression (AoP) This structured approach is essential for accurate and reliable measurements because the AoP and HSD are inherently interrelated tasks The process involves identifying ultrasound planes, followed by segmenting anatomical structures, and finally detecting key landmarks for precise measurement of AoP and HSD Model Architecture and Supervision The SSTN architecture comprises three primary components: a feature enhancement block, a Video Swin Transformer (VST) encoder, and a decoder with ResConvs and UpConvs for optimized feature extraction The video sequence VL is processed through the Video Swin Transformer to generate multi-level feature maps MS at each stage S prior to the patch merging process Supervision is achieved through a weighted summation of loss functions defined as Losstotal = αLCls + βLSeg + γLEndpoints, which balances classification, segmentation, and landmark detection tasks The classification task is prioritized over the segmentation and landmark detection tasks in the order of execution among these tasks Performance Metrics and Results 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 Table 1 presents the performance of the SSTN model in comparison to other methods, achieving a classification accuracy of 82.0% and an F1 score of 81.7% The model achieves a DSC of 0.910 in the segmentation task, representing a 7.

Title and authors: Rosa: The authors are Jie Gan, Zhuonan Liang, Jianan Fan, Lisa McGuire, Caterina Watson, Jacqueline Spurway, Jillian Clarke, and Weidong Cai. They’ve put together this sequential spatial-temporal network framework to tackle the problem of automatically assessing fetal head descent during labor. Dev And what it means for us is that instead of just getting a number without knowing the steps taken, this model lets us see how it arrives at those numbers by breaking down the process into specific stages like classification and segmentation.

Taro: I’m curious about how they handle the temporal aspect here. Since this is an ultrasound video, they have to look at what happens over time, which is something we deal with all the time when you're testing real-world failure modes. Rosa Exactly. The Sequential Spatial-Temporal Network is designed to capture that temporal information across the video sequence to improve accuracy in these labor assessment tasks.

Dev: It’s trying to go beyond just looking at each frame individually, which is where a lot of previous work has hit a wall because they treat the tasks as separate things. Taro So they're suggesting something that connects those stages together sequentially?

The paper's summary: Rosa: Basically, the core idea is to break down the assessment into three stages. First, classification, which decides what’s going on. Then segmentation, where it identifies structures like the pubic symphysis and the fetal head. And finally biometry, which measures things like Head Symphysis Distance and Angle of Progression based on those identified landmarks. Dev So they’re linking all these parts together using a weighted loss function to train the model simultaneously across all those different tasks at once.

Taro: That simultaneous learning part is interesting from a robustness standpoint. If you only trained it for one thing, say just measuring distance, it might fail when the image quality drops or the fetal positioning shifts unexpectedly. Rosa Right, and they use this weighted loss function to balance classification loss, segmentation loss, and landmark detection loss at the same time. It’s a way of making sure that learning to segment doesn't completely overpower learning how to classify what’s happening in the video sequence.

Dev: The paper points out that the tasks are inherently related; you can't really get a good measurement of distance without knowing exactly where those boundaries are in the image, and they address this by making them interdependent during training. Taro So, when we look at the results, it’s not just one number for accuracy; they show how well it does on classification, segmentation, and biometry metrics all together.

The paper's improvements: Rosa: The authors highlight several ways their approach is better than what came before. They point out that previous studies often treated the tasks as independent processes, which they say can lead to suboptimal results when you need them to work together in a clinical setting. Dev And they address the issue of low signal-noise-rate characteristic in ultrasonic samples, saying that this noise isn't something we can just ignore when dealing with these medical images.

Taro: The architecture itself is built around a feature enhancement block, a Video Swin Transformer as the main encoder, and then a decoder using ResConvs and UpConvs for optimized feature extraction. It seems like they’re trying to get the best of both worlds—capturing temporal context with the transformer while keeping spatial information intact through those skip connections in the decoder.

Rosa: They specifically mention that during inference, those UpConvs layers filter frames based on classification results, which means it can use what it learns about the state of labor to decide which parts of the video sequence are most important for calculating measurements. Dev That filtering based on classification seems like a smart way to manage computational load while maintaining relevance for the final biometric assessment.

Taro: The paper also details their composite loss function, which combines Dice score, pixel-wise cross entropy loss, and intersection over union loss for the segmentation part. It’s not just one metric; they are optimizing boundary accuracy and region overlap simultaneously when identifying things like the pubic symphysis or the fetal head.

Conclusion: Rosa: So to wrap up, this Sequential Spatial-Temporal Network for Interpretable Automatic Ultrasonic Assessment of Fetal Head during labor provides a structured way to look at this problem by linking classification, segmentation, and biometry in one interconnected framework. Dev It shows that when you treat these tasks sequentially with a weighted loss function, you get better results than treating them as isolated problems.

Taro: What I take away is that the system's ability to use those skip connections in the decoder to refine segmentation based on detected landmarks is key for ensuring it recovers gracefully when the input data isn't perfect. Rosa And we see concrete gains, like reducing mean absolute error by eighteen percent for Angle of Progression and twenty-two percent for Head Symphysis Distance compared to existing models.

Dev: So, this paper gives us a solid framework that’s both accurate and transparent because you can trace the decision path through those sequential stages. Rosa It’s a big step toward making automated assessment in labor environments more reliable by connecting the temporal flow of the video directly to the precise measurements we need.

More episodes

← Home