In an internal technical validation of intrapartum transperineal ultrasound imaging, an automated segmentation workflow demonstrated close agreement with senior physicians, achieving an Angle of Progression mean absolute error of 5.4293 degrees and a Head-Symphysis Distance mean absolute error of 2.9904 pixels. The research, published in Frontiers in Medicine by investigators from Shandong Province Maternal and Child Health Care Hospital, Jinan Third People’s Hospital, Shandong Sports Rehabilitation Research Center, and Tianjin University of Sport, evaluated whether semantic segmentation of the pubic symphysis and fetal head supports reproducible anatomical measurements during labor.
The diagnostic evaluation focused on two established geometric indicators of labor progress: the Angle of Progression and the Head-Symphysis Distance. To test whether automated image analysis could reliably derive these metrics from intrapartum transperineal ultrasound scans, the research team analyzed image data from the PSFHS dataset. The researchers partitioned this dataset into 951 training images, 136 validation images, and 271 test images. To prevent identical images from contaminating different evaluation stages, exact-image hashes were identified and grouped strictly within their designated splits.
Model development incorporated extensive architectural exploration across the validation framework. The investigators compared nine different segmentation architectures during the development phase before selecting a UNet++ based workflow. Once the UNet++ workflow was selected, its configuration was finalized prior to conducting the formal test evaluation, ensuring that architectural choices remained fixed before analyzing the final holdout test images.
To perform the primary measurement evaluation, the authors implemented an ensembling procedure utilizing three independently trained UNet++ models. For each ultrasound scan within the 271-image test set, the Angle of Progression and Head-Symphysis Distance values generated by the three independent models were averaged. These combined model estimates were then compared directly against a clinical reference benchmark established by calculating the arithmetic mean of independent manual measurements recorded by three senior physicians.
The primary test evaluation demonstrated consistent alignment between the averaged model outputs and the physician consensus. The Angle of Progression calculation yielded a mean absolute error of 5.4293 degrees, with a 95 percent confidence interval ranging from 4.8579 to 6.0505 degrees. For the Head-Symphysis Distance, the workflow achieved a mean absolute error of 2.9904 pixels, with a 95 percent confidence interval between 2.5915 and 3.4285 pixels. Across the entire 271-image test evaluation, the automated system recorded zero geometric measurement failures.
Evaluation of the individual models before averaging confirmed stability across independent training iterations. The individual-model mean absolute errors ranged from 5.7808 to 5.9612 degrees for the Angle of Progression and from 3.1931 to 3.3237 pixels for the Head-Symphysis Distance. Across all training runs, the models achieved a foreground Dice similarity coefficient of 0.9449 with a standard deviation of 0.0011, reflecting precise spatial overlap in delineating the pubic symphysis and fetal head boundaries.
An ablation analysis conducted during the validation phase underscored the distinction between boundary segmentation overlap and clinical calculation accuracy. This validation ablation revealed that a small gain in segmentation overlap could accompany larger geometric measurement errors. This finding demonstrated that optimizing pixel-level overlap alone does not guarantee superior measurement accuracy, confirming the necessity of evaluating the entire end-to-end pipeline from initial image segmentation through to final geometric calculation.
The investigators emphasized that these findings provide an internal technical validation against a physician mean reference rather than a demonstration of clinical efficacy. Critical limitations remain within the current workflow: the Head-Symphysis Distance remains expressed solely in uncalibrated pixels, and patient-level independence could not be verified within the dataset partitions. Consequently, the authors concluded that external validation, reliable physical calibration, and prospective evaluation are essential before the workflow can be established for clinician-led intrapartum assessment.
Subject of Research: Automated measurement of fetal head descent in intrapartum transperineal ultrasound
Article Title: Segmentation-assisted fetal head descent measurement in intrapartum transperineal ultrasound
Article References: Qiu, J., Li, S., Qian, D., Jiang, B., Song, L., Liu, D., & Yang, Z. (2026). Segmentation-assisted fetal head descent measurement in intrapartum transperineal ultrasound. Frontiers in Medicine, 13, Article 1921165. https://doi.org/10.3389/fmed.2026.1921165
Image Credits: AI Generated
DOI: 10.3389/fmed.2026.1921165
Keywords: transperineal ultrasound, intrapartum ultrasound, fetal head descent, angle of progression, head-symphysis distance, pubic symphysis, UNet++, semantic segmentation, PSFHS dataset, mean absolute error, foreground Dice, ensemble modeling, geometric measurement, image hashing, ablation study, medical image segmentation, intrapartum assessment
News Source: Ophelia Keating. (November 21, 2026). Ultrasound Segmentation Workflow Estimates Fetal Head Angle Within Six Degrees. Scienmag.



