Automated Conversion of Ultrasound Pixel Dimensions to Millimeters using Deep Learning Models

Mahmood Alzubaidi, Mowafa S. Househ

Abstract: Ultrasound imaging is a widely used method in prenatal care to obtain fetal biometrics. The automatic conversion of these biometrics from pixels to millimeters (mm) by the ultrasound machine enables physicians to evaluate fetal development. However, the metadata file containing pixel dimensions is often incomplete or missing, presenting a challenge for developing artificial intelligence (AI) applications for fetal ultrasound images. This study proposes a solution that employs pre-trained deep learning models to predict pixel size in mm, thereby automating the labeling process for building AI applications for fetal ultrasound images. The study utilized 2,835 fetal head ultrasound images to train, validate, and test six deep-learning regression models for the conversion of pixels to mm. The evaluation of the deep-learning models involved three steps: traditional evaluation metrics, descriptive analysis, and statistical approach. The results from the three evaluation stages showed that the Xception model outperformed the other models, achieving an R-squared (R2) value of 0.8535 and a mean squared error (MSE) of 0.00028 when predicting pixel size in mm on the test dataset. The descriptive analysis yielded a standard deviation (SD) of 0.0449, while Spearman’s rank correlation coefficient was 0.841.