EP22.15: Feasibility of automated recognition of fetal abdominal circumference on antenatal ultrasound by using convolutional neural network

By using convolutional neural network (CNN), a program was designed to recognise fetal AC plane on ultrasound which takes account of physicians' decision process, anatomical structures, and the characteristics of ultrasound image. We planned to evaluate the feasibility of automated recognition of appropriate AC plane by machine learning using CNN. This study included seventy singleton pregnancies between 20+0∼34+6 weeks of gestation without intra-abdominal anomalies and we analysed 458 AC images. Transverse AC images acquired by three different qualified physicians. Following thirty-five cases of training on recognition of proper AC plane by previously described machine learning program, 458 AC images were reviewed by two certificated obstetricians and were evaluated in adequacy. And inter-observer agreement was analysed by using Cohen's kappa coefficient. There was statistically difference between human and CNN in recognition of appropriate AC plane. Inter-observer agreement between two obstetricians was 94.88% (Kappa value 0.85), however, inter-observer agreements between CNN and each obstetricians were 60.48% (Kappa value 0.11) and 62.45 % (Kappa value 0.15), respectively, There were considerable differences between human and CNN program in recognition of appropriate AC plane. Improvement is needed that fine recognition of fetal abdominal structures which define proper AC plane by feedback into CNN program.