Semantic Segmentation Based on Artificial Intelligence: An Effective Paradigm for Automated Estimation of Fetal Head Circumference in Ultrasound Images

For several decades the obstetricians have measured fetus head circumference (HC) manually from ultrasound images captured from womb in order to estimate its gestational age, size, and weight. Unfortunately, some factors may hamper the performance of this technique including its subjective nature, different diagnoses by different specialists and its time-consuming procedure. Computer Aided Diagnosis has been introduced as an effective paradigm to address the aforementioned problems, but the low contrast of the ultrasonic images, its low signal-to-noise ratio, acoustic shadows, and speckle noise are the limiting parameters of these methods. Thus, in recent years, the use of artificial intelligence has become an inevitable choice to solve this problem due to its ability to construct comprehensive models for the fetus and the background. In this article, a new scheme is presented in artificial intelligence framework in order to estimate fetal HC, which is based on semantic segmentation by using deep learning. In the proposed scheme, the region related to the fetus is separated from the background by using a UNET deep neural network which has been trained by utilizing ultrasound images captured from the womb, in which the fetal head area has been labeled by a specialist. The UNET may promote the performance of HC estimation by applying the sliding window technique which needs fewer images to increase the model performance. Furthermore, such an architecture may promote localization tasks as well as creating distinguished class label for each pixel thanks to create a local patch for each pixel. In order to evaluate the effectiveness of the semantic segmentation scheme it has been implemented as a software package which may test on real HC۱۸ dataset contains a total of ۹۹۹ two-dimensional labeled ultrasound images of the standard plane that can be used to measure the HC. The testbed was prepared on python۳ and TensorFlow paradigms on Intel ® Core i۷-۱۰۷۰۰ computer with Ubuntu ۲۰.۰۴ operation system, ۳۲ GB RAM, and an NVIDIA ۲۰۸۰ Ti. We used Google Colab, a GPU framework made available by Google, to run the program. The results obtained from the above evaluations demonstrated the effectiveness semantic segmentation paradigm in estimation fetal head parameters in such way that the Absolute Difference measure (i.e., ADF) between the ground truth around the fetal head and the estimation of the proposed method has been in range of [۲.۳۵ - ۲.۷۵] millimeters. In the same way, Dice's parameter for the proposed method has been obtained in range of [۹۶.۷۵ – ۹۷.۷] percent and Jaccard's parameter has been in range of [۹۴- ۹۵] percent. These parameters indicate the acceptable similarity of the head circumference boundary extracted from the proposed method to the actual fetal head circumference boundary. The low difference between actual and estimated HC of fetus (e.g., the ADF) as well as the high similarity between real and estimated borders (e.g., Dice and Jaccard measures) showed that the use of deep learning based semantic segmentation can be developed and utilized as an option with the potential of fetal head circumference estimation in practical applications.

کلیدواژه ها:

Fetal Head Circumference ، Artificial Intelligence ، Semantic Segmentation ، Deep Learning ، UNET.