臨床超音波測量數據運用類神經網路建立預測胎兒體重之模型

Using artificial neural network model with clinical ultrasound measurements for fetal weight prediction

白詩涵(Shih-Han Pai)

指導教授 ： 邱泓文

臺北醫學大學/醫學科技學院/醫學資訊研究所/碩士(2012年)

https://doi.org/10.6831/TMU.2012.00138

摘要

目前臨床上主要仍是以傳統回歸公式分析方式來預估胎兒體重，主要以超音波數據BPD (Biparietal Diameter；雙頂骨徑)、OFD(Occipital frontal diameter；枕額徑)、AC (abdominal circumference；胎兒腹圍)、FL(Femur length；股骨長度)作為評估的參數，仍會有所謂的評估上的誤差，超音波預估體重誤差率大概為10~15%，甚至提高至25%。然而，體重的預估對於胎兒的評估是不可或缺的，也因此減少胎兒體重上的預估誤差，對於在產前的判斷是十分重要。本研究目的是利用2D超音波數據，運用ANN分析胎兒出生體重，來改善傳統回歸分析公式所產生的誤差值。總共有13個輸入的變項：BPD、OFD、FL、HC (Head circumference；頭圍)和AC、STT (Soft-tissue thickness；軟組織的厚度)、DTB (Days to birth；天數差異)、子宮底高度 (cm)、母親腹圍 (cm)、母親體重 (kg)、母親身高 (cm)，輸出參數：娩出體重 (gm)。其中STT指的是胎兒大腿骨中段的皮下脂肪肌肉厚度，以及DTB是指天數的差異。因無法確實掌握超音波測量的時間與胎兒出生時間的距離故將天數納入為變數之一。在訓練組方面，我們利用70位胎兒完成超音波測量，輸入參數有：BPD (mm)、OFD (mm)、FL (mm)、HC (mm)和AC (mm)、STT (mm)、DTB (Days)、子宮底高度 (cm)、母親腹圍 (cm)、母親體重 (kg)、母親身高(cm)，輸出參數：娩出體重 (gm)。驗證組方面，有35位胎兒完成超音波測量，輸入參數有：BPD (mm)、OFD (mm)、FL (mm)、HC (mm)和AC (mm)、STT (mm)、DTB (Days)、子宮底高度(cm)、母親腹圍 (cm)、母親體重 (kg)、母親身高 (cm)，在此篇研究中，利用ANN中的多層式感知器 (multi-layer perception, MLP)架構作為系統的模型，共分為五組模型。結果顯示，Model 1、Model 2在驗證組的平均絕對誤差和平均絕對百分比誤差之結果比傳統公式的結果差。然而，在Model 3、Model 4和Model 5驗證組的平均絕對誤差和平均絕對百分比誤差之結果，分別為246.307gm，7.483%、213.656 gm，6.491%和251.380gm，7.637%。這三組模型，優於傳統公式驗證組的平均絕對誤差和平均絕對百分比誤差之結果，分別為341.257gm和10.368 %。因此，可以利用ANN模型，作為胎兒體重預估的新工具，並且加入STT或結合DTB的超音波測量參數，可以提供新的胎兒出生體重評估方式。

關鍵字

類神經網路 ； 胎兒體重評估 ； 超音波

並列摘要

Traditionally, clinical estimation of fetal weight is based on regression formula with ultrasound data of biparietal diameter (BPD), occipitofrontal diameter (OFD), abdominal circumference (AC), femur length (FL) as the parameters for assessment. Such method is validated with about 10 to 15 percent, and even to 25 percent of error rate. However, the weight estimation for fetal assessment is necessary, thus reducing the error of fetal weight estimation for prenatal judgment is very important. The purpose of this study is to train and validate an Artificial Neural Network (ANN) model with clinical ultrasound measurement data for birth of fetal weight prediction and expect it to improve the estimation results. In addition to BPD, OFD,AC and FL that have been used in traditional formula, head circumference (HC), uterine fundus of height degree (cm),mother abdominal circumference (cm)、mother body weight (kg),mother height (cm),soft-tissue thickness (STT) and days to birth (DTB) were added as the input variables of ANN. Herein the STT is thickness of the subcutaneous fat and muscle of the middle section of the fetal femur, and DTB is the time difference between birth and ultrasound measurement time that could have the impact to weight estimation. We collected the 70 fetuses as training group and 35 fetuses as testing group. The multi-layer perceptron (MLP) ANN model with input parameters BPD (mm) , OFD (mm) , HC (cm), AC (mm), FL (mm) , uterine fundus of height degree (cm), mother abdominal circumference (cm), mother body weight (kg), mother height (cm),STT (mm), DTB (day) and an output parameter, birth weight (g) was trained for five works. In results, Model 1and Model 2 of the mean absolute error (MAE) for birth weight estimation and mean absolute percent error (MAPE) were poor than traditional formula, in the testing group. However , the results of Model 3,Model 4and Model 5, the mean absolute error (MAE) for birth weight estimation and mean absolute percent error (MAPE), respectively of 246.307gm, 7.843% ; 213.656gm ,6.491% and 251.380gm ,7.637% ,in the testing group. These three models was better than the traditional formula in the testing group, MAE and MAPE is 341.257gm and 10.368 % respectively. Therefore, the ANN model can be used as a new tool for fetal weight estimation, and our new model based on ANN with new parameters STT and DTB are able to improve the fetal weight assessment.

並列關鍵字

Artificial neural network ； Fetal weight estimation ； Ultrasound

參考文獻

Blumenfeld, Y. J., Lee, H. C., Pullen, K. M., Wong, A. E., Pettit, K., & Taslimi, M. M. (2010). Ultrasound estimation of fetal weight in small for gestational age pregnancies. J Matern Fetal Neonatal Med, 23(8), 790-793.

Boulet, S. L., Alexander, G. R., Salihu, H. M., & Pass, M. (2003). Macrosomic births in the united states: determinants, outcomes, and proposed grades of risk. Am J Obstet Gynecol, 188(5), 1372-1378.

Chuang, L., Hwang, J. Y., Chang, C. H., Yu, C. H., & Chang, F. M. (2002). Ultrasound estimation of fetal weight with the use of computerized artificial neural network model. Ultrasound in Medicine and Biology, 28(8), 991-996.

Farah N, S. B., Donnelly V, Rafferty G, Turner M. (2009 ). What is the value of ultrasound soft tissue measurements in the prediction of abnormal fetal growth? American Journal of Obstetrics & Gynecology 29(6).

Farmer, R. M., Medearis, A. L., Hirata, G. I., & Platt, L. D. (1992). The use of a neural network for the ultrasonographic estimation of fetal weight in the macrosomic fetus. Am J Obstet Gynecol, 166(5), 1467-1472.