AI Reads Uterine Ultrasound Scans to Measure Endometrium in Seconds

An artificial intelligence system can now identify the endometrium on uterine ultrasound images, measure its thickness, and classify its pattern automatically, performing all three tasks in roughly 2.2 seconds per image. The tool, described in a study published in Discover Imaging, was developed by researchers working with Canadian fertility clinics and trained on nearly 6,000 annotated ultrasound images from patients undergoing standard in vitro fertilization care. Its reported accuracy exceeds 90 percent across segmentation, measurement, and pattern classification, marking what the authors describe as the first automated method for assigning endometrial pattern on ultrasound.

The endometrium, the tissue layer lining the uterine cavity, is routinely measured before embryo transfer in medically assisted reproduction cycles. Clinics commonly use a thickness cutoff of around 7 millimeters to decide whether to proceed with transfer, and they also assess whether the endometrium shows a trilaminar, or triple-line, appearance or a homogenous one. Yet these assessments are subject to broad interpretation and significant error, with variability in imaging method, caliper placement, image interpretation, and potential human bias around target thickness cutoffs. The clinical utility of thickness and pattern assessment is itself debated in the literature, but the measurements remain widely used and unlikely to be abandoned, which motivated the researchers to ask how accuracy, reliability, and speed could be improved.

Automating endometrial analysis is harder than automating other reproductive imaging tasks. Existing artificial intelligence applications for medical image interpretation tend to rely on high-contrast features, such as the dark polygons of follicular fluid against lighter ovarian stroma, or isolated embryos against translucent media. Endometrial and myometrial tissues have similar densities, producing acoustic images where the layers can be resolved but only at low contrast. The endometrium also changes continuously across the menstrual cycle, cycling through proliferation, differentiation, secretion, and degradation, and is composed of two layers, the stratum basalis and the stratum functionalis, that can be difficult to distinguish from each other and from surrounding myometrium depending on cycle stage.

To build the system, the team accessed de-identified patient records dated between 2014 and 2023 from Canadian fertility clinics. Transvaginal ultrasound images were captured in a standardized way: the probe was positioned to identify the thickest aspect of the endometrial echoes roughly 5 to 10 millimeters from the myometrial-endometrial junction at the uterine fundus in the mid-sagittal plane, then rotated 90 degrees to capture transverse images at the level of the uterine cornu. The dataset comprised 5,985 unique ultrasound images from 5,985 individual patient cases, each annotated by trained personnel using purpose-built software, with every annotation cross-validated by a second annotator and disagreements resolved by the most experienced senior individual.

The images were randomly divided into training, validation, and test sets at a 70 percent, 10 percent, 20 percent ratio, yielding 4,189 training images, 598 validation images, and 1,198 test images. The researchers used the YOLO version 8 framework, a single-stage object detection model chosen for its combination of high accuracy and high inference speed, which also makes deployment on non-specialized clinical hardware feasible. They considered alternatives: U-Net architectures offer comparable segmentation performance, transformer models demand larger datasets, and Meta’s Segment Anything Model required substantial manual selection and failed on the grayscale, fuzzy ultrasound images. For thickness measurement, the team adapted a principal component analysis approach, treating the segmented endometrium as a slightly deformed ellipse and computing its minor axis to locate the thickest point perpendicular to the uterine lumen.

The segmentation results were strong. On the test set, the main endometrial segmentation model scored 97.7 percent accuracy, 98.4 percent precision, and 97.3 percent sensitivity at the standard 50 percent mask-overlap threshold. A full 92.8 percent of test predictions reached an intersection-over-union value greater than 0.75, the study’s high-performance category, and the mean IoU was 0.91. The confusion matrix showed accuracy of 0.98, precision of 0.99, and a Matthews correlation coefficient of 0.98. Because no images lacking endometrium were included, there were no true negatives in the dataset, making specificity undefined for the object detection task.

Thickness measurement achieved a mean absolute error of 0.89 millimeters in length, 2.81 millimeters in the position where the measurement was placed along the axis perpendicular to the lumen, and 5.5 degrees in orientation relative to the lumen. That sub-millimeter length error compares favorably with previously published automated methods, which showed measurement errors of up to 4 millimeters on much smaller datasets. The authors note that with a 0.89 millimeter error, any endometrium thicker than 7.9 millimeters or thinner than 6.11 millimeters would be correctly classified against the common 7 millimeter transfer threshold; only cycles falling between those bounds, fewer than 20 percent in their experience, could potentially be misassigned and flagged for human review. Human measurement error, by comparison, is estimated at 10 to 20 percent of total thickness.

Pattern classification was the novel component. The model detected the presence or absence of a visually distinct stratum functionalis layer, using its presence as the indicator of a trilaminar pattern and its absence as a homogenous pattern. Because only 110 fully annotated functionalis images were available, the team generated synthetic functionalis polygons by shrinking the endometrial polygons to approximate the layer’s thickness. The functionalis detection model itself achieved 83.1 percent accuracy, 82.0 percent precision, and 80.5 percent recall, but as a binary classifier sorting patterns into homogenous versus triple-line, it reached 91.9 percent accuracy, correctly predicting pattern on 1,101 of the 1,198 test images. A more traditional VGG image classifier tested for comparison performed worse, at 75 percent.

The cohort itself reflected routine fertility practice: mean patient age was 36.49 years, frozen-thaw embryo transfer cycles made up 85 percent of cases, mean endometrial thickness was 9.22 millimeters, and homogenous and triple-line patterns were nearly evenly split at 48 and 52 percent. Notably, about 6.5 percent of the images, 388 of 5,985, carried notations for uterine fibroids, many of which cast shadows or deformed the uterine cavity, suggesting the method holds up under real-world imaging conditions. The authors caution, however, that excessive motion artifact, imbalanced image gain, over- or under-saturated pixel densities, and anatomic shadowing from fibroids or fluid can still significantly degrade performance.

Several limitations temper immediate clinical adoption. The training images were acquired outside the standard imaging timing used in many clinics, roughly 48 hours before a day-5 or day-6 embryo transfer, and in the transverse rather than the mid-sagittal plane that standard practice typically relies upon, although early modeling suggested positive performance on mid-sagittal images as well. The utility of the tool also presupposes the utility of thickness and pattern assessment itself, which remains debated. The researchers relied on a single segmentation technique and plan to explore alternatives such as U-Net, DeepLabv3+, and SAM, along with larger validation datasets. They emphasize that the system is intended to reduce clinical workload and inter-observer variability through standardization, not to replace clinical judgment: the final decision on whether a patient proceeds with embryo transfer remains with the clinician.

Subject of Research: An AI system for automated endometrial segmentation, thickness measurement, and pattern classification on uterine ultrasound images

Article Title: Automated endometrial segmentation, thickness measurement and pattern prediction on uterine ultrasound images

Article References: Pierson, H. E., Shand, Z., Invik, J., Wang, Y., Ohal, V., Smith, K., Patel, D., & Pierson, R. A. (2025). Automated endometrial segmentation, thickness measurement and pattern prediction on uterine ultrasound images. Discover Imaging, 2(1), Article 15. https://doi.org/10.1007/s44352-025-00019-w

Image Credits: AI Generated

DOI: 10.1007/s44352-025-00019-w

Keywords: artificial intelligence, ultrasound, endometrium, in vitro fertilization, medical imaging, machine learning, YOLOv8, image segmentation, endometrial thickness, reproductive medicine, computer vision, embryo transfer