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

To develop an AI-based model for automated identification, thickness measurement, and pattern classification of endometrial tissues in uterine ultrasound images. We utilized a multi-stage model building and training approach with different methods for segmentation, thickness measurement, and pattern assignment. 5985 unique annotated ultrasound images were utilized in model training and validation. The annotated images were randomly divided into three groups (training data, validation data, and test data) at a 70%/10%/20% ratio. 4787 images were used to develop the model; 1198 images were used to evaluate performance. Identification of the endometrium was done utilizing direct object detection and instance segmentation systems with single stage object detectors (YOLO v8). A principal component analysis approach was adapted to determine endometrial thickness. We also trained computational models to assign an endometrial pattern of either ‘trilaminar’ or ‘homogenous’. Intersection over union (IoU), confusion matrices, and error calculations were conducted to assess the model’s proficiency and accuracy. The endometrial segmentation model performed at 98% accuracy on the confusion matrix and 92.8% of test data had an intersection over union (IoU) value > 0.75. The endometrial thickness measurement achieved mean absolute error rates of: ± 0.89 mm in length, ± 2.81 mm in position (along the perpendicular-to-lumen axis), and 5.5-degrees in orientation relative to the lumen. Automatic assignment of pattern as either ‘trilaminar’ or ‘homogenous’ achieved a 92% accuracy rate. A novel automated method for routine endometrial thickness and pattern assessment is demonstrated. We report a first-of-its kind method for automated endometrial pattern assignment (homogenous / triple-line) on uterine ultrasound images.