همایش ملی بیوانفورماتیک ایران

صفحه اصلی / 4th international edition and 13th Iranian Conference on Bioinformatics

Fully Convolutional Neural Networks for Volumetric Segmentation of Ultrasound Images: An Effective Tool for Automated Estimation of Fetal Head Circumference

نویسندگان :

Seyed Vahab Shojaedini¹ Mohammad Momenian²

1- سازمان پژوهش های علمی و صنعتی ایران (IROST) 2- دانشگاه آزاد اسلامی واحد الکترونیکی

کلمات کلیدی :

Fetal Head Circumference،Artificial Intelligence،Volumetric Segmentation،Deep Learning،VNET

چکیده :

While manual calculations remain a straightforward method for estimating fetal Head Circumference (HC) from sonographic images (Zhao et al., 2022), several factors can hinder its effectiveness. These include the subjective nature of image interpretation, variability among experts’ assessments, and the labor-intensive nature of the process. Consequently, the integration of artificial intelligence (AI) has emerged as a critical solution, offering the ability to generate detailed representations of the fetus and its surrounding environment. Despite the adoption of various deep neural networks in this domain, challenges such as low ultrasound image clarity and a limited signal-to-noise ratio persist (Lee et al., 2023). This research introduces an innovative AI-driven approach for estimating HC based on deep learning volumetric segmentation. The method employs a tailored 2D VNET architecture designed to isolate the fetal region from the background (Qian et al., 2024). By incorporating a sliding window strategy, this architecture enhances the accuracy of HC estimation with a smaller number of images, improving model performance. Additionally, its design supports precise localization tasks and assigns unique class labels to each pixel by generating localized patches for individual pixels. To evaluate the effectiveness of the proposed method, it was implemented as a software package and tested on the HC18 dataset, which consists of 999 labeled sonographic images of the standard plane used for HC measurement. The experiments were conducted using Python 3 and TensorFlow on a system equipped with an Intel® Core i7-10700 processor, 32 GB of RAM, an NVIDIA 2080 Ti GPU, and the Ubuntu 20.04 operating system. Google Colab’s GPU platform was also utilized to execute the program. The results demonstrated the efficacy of the volumetric segmentation model in estimating HC parameters. The Absolute Difference (ADF) between the actual and predicted values for HC was obtained averagely equal to 2.89 millimeters. Additionally, the proposed method achieved a Dice coefficient of [95.22±2.16] percent and a Jaccard index of [93.7±2.4] percent. These metrics highlight the minimal discrepancy between the real and predicted HC measurements and the high degree of similarity between true and estimated contours. Consequently, deep learning-based volumetric segmentation holds significant promise as a practical tool for HC estimation in clinical settings.