Intrapartum Ultrasound Video Analysis for Enhanced Labor Decision Making
Al Nasi, Salem Muhsin Ali Binqahal
Al Nasi, Salem Muhsin Ali Binqahal
Supervisor
Department
Machine Learning
Embargo End Date
2025-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This thesis addresses critical challenges in intrapartum ultrasound (ITU), focusing on automated analysis of fetal head (FH) and pubic symphysis (PS) to facilitate crucial clinical measurements such as the angle of progression (AoP) and head-symphysis distance (HSD). In modern obstetrics, accurate assessment of these parameters is vital for effectively monitoring labor progression. However, existing workflows rely heavily on manual annotation and measurement, which are time consuming, prone to human error, and susceptible to label inconsistencies, particularly in video-based data.
Problem Statement and Hypothesis. Ultrasound video datasets commonly exhibit label noise at the frame level, where entire clips are annotated as “positive” or “negative” despite frame-to-frame variability. Mislabeling diminishes model reliability, leading to suboptimal classification, segmentation, and biometric estimation. We hypothesize that a multitask framework, enriched with robust noise-handling mechanisms such as testtime augmentation (TTA), confident learning, and curriculum learning, can overcome these noisy labeling barriers and improve performance across classification, segmentation, and clinical measurement tasks.
Methodology. The research is divided into two major components. First, we develop a multitask learning (MTL) pipeline capable of simultaneously classifying standard-plane frames and segmenting FH/PS structures. This pipeline integrates pseudolabeling and weak supervision techniques to mitigate the scarcity of fully annotated data. We also introduce dynamic loss scaling and use a schedule-free optimizer to accommodate the differing learning dynamics of classification and segmentation tasks. Second, we extend these methods to a noise robust framework for medical videos by incorporating Confident Learning (CL) and Test Time Augmentation (TTA). The CL approach prunes and ranks potentially mislabeled frames and videos, while TTA improves framelevel confidence estimation through simulated probe movements. Finally, a curriculum learning schedule gradually introduces uncertain or challenging samples, mirroring the way clinicians learn to interpret complex ultrasound clips only after mastering unambiguous examples.
Findings and Contributions. Experiments on multiple intrapartum ultrasound datasets and challenge benchmarks (e.g., IUGC 2024) confirm notable advancements in classification accuracy, segmentation Dice, and biometric reliability. Specifically, the framework achieves a twofold improvement in detecting mislabeled frames, reduces average label noise by over 4%, and lowers AoP error by over 1.0–2.0 degrees compared to traditional baselines. Moreover, restricting TTA to anatomically variable structures (like the PS) further optimizes the tradeoff between computational overhead and performance gains.
Benefits. This thesis demonstrates that data-centric learning strategies, when combined with targeted augmentations and curriculum-based training, substantially improve model robustness under noisy conditions. From a clinical standpoint, enhanced Ao P and HSD measurements have the potential to improve labor management, decrease unnecessary interventions, and facilitate broader adoption of automated intrapartum ultrasound solutions. Taken together, these findings advance the stateoftheart in obstetric imaging analysis, offering a scalable blueprint for addressing noisy labels in medical video datasets more broadly.
Citation
Salem Muhsin Ali Binqahal Al Nasi, “Intrapartum Ultrasound Video Analysis for Enhanced Labor Decision Making,” Master of Science thesis, Machine Learning, MBZUAI, 2025.
Source
Conference
Keywords
Intrapartum Ultrasound, Multitask Learning, Confident Learning, Curriculum Learning, Biometric Measurements, Pubic Symphysis