Anatomical Object Detection in Fetal Ultrasound: Computer-Expert Agreements

Abstract

An alternative approach to the sliding window method for the detection of anatomical objects in fetal ultrasound image is proposed in this pper. The global feature symmetry map derived from the local phase computation of the images was integrated within a machine learning framework that trains a local classifier using local Haar features from intensity images. This provides a computationally cheap step before invoking a local object detector to be applied in plausible locations. The proposed method exhibits better generalization capability when tested on 2384 images with an accuracy of 82.75% and 72.55% for the detection of the stomach and the umbilical vein, respectively. It also has faster computation time than the typical local object detector with the sliding window approach. It was observed that the method achieved high accuracy detection by focusing only on the high probability region and discarding many false positives candidates as in the sliding-window method. The agreement between the automated method and the experts in detecting the presence and absence of the stomach and the umbilical vein were also compared. The results indicate that the agreement between the automated method and the experts were very good for computer-random-selected images and the agreement was comparable to inter-experts agreement. © Springer-Verlag Berlin Heidelberg 2014.