<p>Ultrasound is the most preferred modality for prenatal screening because of its low cost and easy availability. The fetal head circumference measurement is a useful biometric to estimate the gestational age and due date and to detect growth abnormalities in the fetus. The existing methods are time-consuming, challenging and need skilled sonographers because of a lot of inter and intra-operator variability. We proposed a deep learning-based solution that learns multi-scale and self-attention based features to automatically segment the fetal head from 2D ultrasound images. The multi-scale and multi-feature modules learn robust features. Furthermore, the attention module incorporated in the network increases the network performance by enhancing the relevant features and suppressing the irrelevant ones. The network is evaluated on the benchmark HC18 Grand Challenge data set containing ultrasound images of all three trimesters without any growth abnormality. In comparison to UNet, our proposed method has shown better performance, i.e. i) 0.5 % increase in dice coefficient and ii) 24% decrease in mean absolute difference, iii) Three times lesser parameters and takes half the time to train. The network has shown outstanding performance and is superior to state-of-the-art methods. It can be used in real-time since the inference time is lesser than that of a trained sonographer.</p>