We developed an artificial intelligence (AI) method for estimating fetal weights of Japanese fetuses based on the gestational weeks and the bi-parietal diameter, abdominal circumference, and femur length. The AI comprised of neural network architecture was trained by deep learning with a dataset that consists of ± 2 standard devia-tion (SD), ± 1.5SD, and ± 0SD categories of the approved standard values of ultrasonic measurements of the fetal weights of Japanese fetuses (Japan Society of Ultrasonics in Medicine [JSUM] data). We investigated the residuals and compared 2 other regression formulae for estimating the fetal weights of Japanese fetuses by t-test and Bland-Altman analyses, respectively. The residuals of the AI for the test dataset that was 12.5% of the JSUM data were 6.4 ± 2.6, -3.8 ± 8.6, and -0.32 ± 6.3 (g) at -2SD, +2SD, and all categories, respectively. The residu-als of another AI method created with all of the JSUM data, of which 20% were randomized validation data, were -1.5 ± 9.4, -2.5 ± 7.3, and -1.1 ± 6.7 (g) for -2SD, +2SD, and all categories, respectively. The residuals of this AI were not different from zero, whereas those of the published formulae differed from zero. Though vali-dation is required, the AI demonstrated potential for generating fetal weights accurately, especially for extreme fetal weights.
Keywords: artificial intelligence; deep learning; fetal weight; neural network; ultrasound biometry.