Learning evaluation of ultrasound image segmentation using combined measures

Objective evaluation of medical image segmentation is one of the important steps for proving its validity and clinical applicability. Although there are many researches presenting segmentation methods on medical image, while with few studying the evaluation methods on their results, this paper presents a learning evaluation method with combined measures to make it as close as possible to the clinicians’ judgment. This evaluation method is more quantitative and precise for the clinical diagnose. In our experiment, the same data sets include 120 segmentation results of lumen-intima boundary (LIB) and media-adventitia boundary (MAB) of carotid ultrasound images respectively. And the 15 measures of goodness method and discrepancy method are used to evaluate the different segmentation results alone. Furthermore, the experimental results showed that compared with the discrepancy method, the accuracy with the measures of goodness method is poor. Then, by combining with the measures of two methods, the average accuracy and the area under the receiver operating characteristic (ROC) curve of 2 segmentation groups are higher than 93% and 0.9 respectively. And the results of MAB are better than LIB, which proved that this novel method can effectively evaluate the segmentation results. Moreover, it lays the foundation for the non-supervised segmentation evaluation system.