A Comparison of Local Descriptors on Cardiac Ultrasound Images

In the literature of pattern recognition and computer vision, local descriptors have been widely used in applications such as shape matching and object recognition. Numerous descriptors have been proposed and evaluated, but little work is reported in the area of medical image, especially ultrasonic images. In this paper, we assess the performance of different local descriptors to detect specific objects in cardiac ultrasound image. Ultrasonic images are particular noisy. It is yet to be determined whether the descriptors from general vision problems can still have ideal performance on ultrasound images. We compare several descriptors such as context, histogram, moment invariant, texture and generic Fourier descriptor (GFD), and use recall and 1-precision to evaluate their performance. Experiments show that moment invariant and GFD have higher recall, but high 1-precision as well. Combination of different descriptors are also evaluated and they turn out to be more effective than single descriptors.