Discrimination of Probe-Stabilized Scene from Neonatal Cranial Ultrasonic Movies by Supervised Learning for Selective Detection of Pulsatile Tissues

抄録

A supervised learning method has been proposed for discriminating probe-stabilized scenes from the neonatal cranial ultrasonic (US) movies to achieve the motion-artifact-free detection of pulsatile tissues. The pulsatile tissues, echogenic tissues with periodic vibrations, were detected in previous studies and their pediatric usefulness was demonstrated in the diagnosis of ischemia and hemorrhage. However, the probe-stabilized scenes should be manually selected to avoid the effect of probe sway before proceeding further analysis with a large US movie archive. In our approach, the frequency spectrum at each pixel was calculated over every 64 frames of the movie and collected to evaluate three feature sets. Then, several models were trained with combined feature sets and four different algorithms including random forest, gradient descent, k-nearest neighbors, and logistic regression. Finally, each model was evaluated with 10-folds cross-validation. The random forest model with combined features of spatially-averaged power spectrum and its difference rate achieved the best performance with maximum sensitivity at zero False-Positive-Rate of 0.931. It represents that more than 90% of the probe-stabilized scene can be automatically extracted without false positives, which enables us to effectively examine the diagnostic performance of pulsatile tissues in early detection of Neonatal Hypoxic-Ischemic Encephalopathy and Intraventricular Hemorrhage.