Attention mechanism optimized neural network for automatic measurement of fetal anterior-neck-lower-jaw angle in nuchal translucency tests

Nuchal translucency (NT) examination has become a standard item in early pregnancy tests because of its clinical value in detecting early gestational fetal abnormalities. Meanwhile, practitioners may sometimes have difficulty meeting the required criteria for NT measurement owing to their complexity or the exorbitant demand for NT tests. Thus, NT image quality control is critical, particularly for the frequently neglected criterion that the fetal head posture (FHP) shall be neither hyperextended nor hyperflexionated. The Nuchal Translucency Quality Review Program (NTQR) defines all FHPs quantitatively using the anterior-neck-lower-jaw angle (ANLJA), namely the angle between the fetal anterior neck and its lower jaw. According to NTQR’s definitions, hyperflexion is defined as ANLJA close to 0 ∘ ; hyperextension, ANLJA greater than 90 ∘ ; and normal FHP, ANLJA between 0 ∘ and 90 ∘ . Focusing on FHP classification, we proposed a novel algorithm combining an attention mechanism and a convolutional neural network to predict ANLJAs. The new approach integrated channel and spatial attention and was capable of swiftly locating regions of interest. It abstracted important image features based on the information weight and attenuated useless information. The manual method was traditionally regarded as the gold standard for ANLJA measurement. Our findings indicated that the proposed algorithm performed admirably in the ANLJA prediction. It surpassed all other comparison algorithms in terms of efficiency.