Cervical texture analysis in mid-trimester scan to predict spontaneous preterm birth: A case-control study

To compare the performance of cervical texture analysis with sonographic cervical length when predicting risk of preterm birth in mid-trimester.

A single-centre, double-blinded, case-control study was performed including singleton pregnancies with spontaneous vaginal delivery. Cases were considered women with preterm delivery before 37 weeks of gestation and controls were those with term delivery at 37 weeks or more. Ultrasound transvaginal images of the cervix obtained at routine mid-trimester scan were retrospectively analysed. For texture analysis, a specific software based on first and second order statistics was developed. Machine learning algorithms and a convolutional neural network (CNN) were constructed to extract texture features. Several feature selection algorithms were used to identify the most relevant ones.

A total of 381 women were included, 178 (46.72 %) cases and 203 (53.28 %) controls. For cervical length ≤25 mm at midterm the diagnostic accuracy in the prediction of preterm birth was 55.91 %, with a sensitivity and specificity of 10.67 % [95 % CI, 6.6-16.2] and 95.56 % [95 % CI, 91.8-98.0], respectively. For texture analysis, the random forest classifier had an accuracy of 77 %, with a sensitivity of 70.1 % [95 % CI, 63.0-76.3] and specificity of 69.7 % [95 % CI, 63.1-75.6]. Similarly, for CNN-based features the accuracy was 75 %, with a sensitivity of 70.6 % [95 % CI, 63.5-76.8] and specificity of 73.2 % [95 % CI, 66.8-78.8].

Cervical texture analysis from ultrasound images in mid-trimester may represent a promising adjunct for identifying pregnant women at risk of spontaneous preterm birth.