AUTOMATED LOCALIZATION OF AMNIOTIC FLUID POCKETS IN OBSTETRIC ULTRASOUND IMAGES USING YOLOV8: A DOCUMENTED DETECTION BASELINE

Abstract

Amniotic fluid volume (AFV) is a routine, non-invasive indicator of fetal well-being, but conventional ultrasound assessment depends strongly on sonographer skill and pocket selection. In this work, we report a deep-learning baseline for localizing visible amniotic-fluid pockets in two-dimensional obstetric ultrasound images. The primary public dataset comprised 3,335 images: 1,570 from the Zenodo Fetal Planes Database and 1,765 from a Roboflow fluid-detection dataset. A further approximately 120 locally collected clinical images were also available, intended primarily for external robustness assessment rather than for training. The 3,335 public images were partitioned using a 70:20:10 split into 2,331 training images, 667 validation images, and 337 independent test images, and a YOLOv8s model was trained on 640 × 640-pixel inputs with a batch size of 16. Validation precision reached 0.9973, recall 1.0000, mAP50 0.9950, and mAP50–95 0.7678. The model output is fluid-pocket localization only: pixel-level segmentation, calibrated Single Deepest Pocket measurement, and clinical classification were not established as validated outcomes in the available experimental record. Accordingly, the contribution here is a documented amniotic-fluid localization baseline built on the full 3,335-image public dataset, with a clear path toward independent testing, segmentation, and clinical measurement, rather than a complete automated AFV-assessment system.