Machine Learning Algorithms for Classification of First-Trimester Fetal Brain Ultrasound Images.

OBJECTIVE

To evaluate the feasibility of machine learning (ML) tools for segmenting and classifying first-trimester fetal brain ultrasound images.

METHODS

Two image segmentation methods processed high-resolution fetal brain images obtained during the nuchal translucency scan: "Statistical Region Merging" (SRM) and "Trainable Weka Segmentation" (TWS), with training and testing sets in the latter. Measurement of the fetal cerebral cortex in original and processed images served to evaluate the performance of the algorithms. Mean absolute percentage error (MAPE) was used as an accuracy index of the segmentation processing.

RESULTS

The SRM plugin revealed a total MAPE of 1.71% ± 1.62 SD (standard deviation) and a MAPE of 1.4% ± 1.32 SD and 2.72% ± 2.21 SD for the normal and increased NT groups, respectively. The TWS plugin displayed a MAPE of 1.71% ± 0.59 SD (testing set). There were no significant differences between the training and testing sets after 5-fold cross-validation. The images obtained from normal NT fetuses and increased NT fetuses revealed a MAPE of 1.52% ± 1.02 SD and 2.63% ± 1.98 SD.

CONCLUSIONS

Our study demonstrates the feasibility of using ML algorithms to classify first-trimester fetal brain ultrasound images and lay the foundation for earlier diagnosis of fetal brain abnormalities.