Publication

Abstract : Ensuring the healthy development of a fetus involves reliable and timely identification of various maternal–fetal anatomical structures during the pregnancy. However, manual screening is highly subjective, laborious, and depends on the knowledge of sonographers in fetal anatomy. To enhance workflow efficiency and improve diagnosis accuracy through automated fashion, a deep learning (DL)-based image classification architecture is proposed in this study for categorizing commonly followed maternal–fetal anatomical structures in 2-D fetal ultrasound (US) images. Automated DL framework utilizes a pre-trained deep convolutional neural network (CNN) and adaptive channel weighting (ACW) for classifying the fetal anatomical structures into various classes, such as the abdomen, femur, thorax, maternal cervix, brain, and other less frequently used planes as other class. Initially, the widely employed pre-trained deep CNN model, VGG-19, extracts anatomical features from the fetal US images. Subsequently, robust discriminative cues are generated by including ACW to mitigate the low inter-class variance of the fetal US images, leading to incorrect assessments that lower classification accuracy. The ACW employs a multiscale channel attention module (MSCAM) at higher-level convolutional blocks of the backbone CNN model to recalibrate the final feature map channel weights using local and global context information. The efficacy of the proposed method was evaluated using two publicly available maternal–fetal US image datasets. Experimental results demonstrate that the proposed method outperforms the recent deep CNN and other competing models by achieving 95.33% and 98.20% classification accuracy on both fetal US image datasets.