Four-chamber plane detection in cardiac ultrasound images based on improved imbalanced AdaBoost algorithm

An improved imbalanced AdaBoost algorithm called ImAdaBoost was proposed for four-chamber plane detection in cardiac ultrasound images. By assigning dynamic weight adjustment factor to each sample according the precision and true positive rate of weak learner, imbalanced classification problem was converted to cost-sensitive classification problem. Template matching method was used to detect the degree of probe rotation, planes with non-zero degree was discarded to reduce the imbalance. Using multi-instance learning method, the ultrasound image was considered as a bag which was composed of small regions, each region was described by SIFT features, the Bag of Words method was used to map instance features to bag features. ImAdaBoost was used to detect four-chamber planes in cardiac ultrasound images. The experimental results indicate that ImAdaBoost is effective on clinical cardiac ultrasound images. True positive rate is increased while false positive rate is controlled within an acceptable range.