Due to the increasing number of medical images being used in medical praxis, the manual processing of the data becomes impossible especially with the vast number of data available, which makes the automatic processing necessary. One of the cases, where the full automation of measuring process is needed is the measurement of static and dynamic parameters of arteries, such as lumen diameter (LD), artery stiffness (AS), or intima media thickness (IMT). The results of such measurements can be used for the prediction of patient’s risk of cardio-vascular events. The measurement process typically consists of two subsequent steps – the detection of the artery (localization in an ultrasound image) and later segmentation of artery wall. Recently, the issue of segmentation (the second step) is quite well explored and this step is fully automated. But still many measurement systems are not fully automated; often the localization step requires the interaction with operator. To achieve the fully automatic system it is required to automate the localization of the artery in order to make it ready for next processing.
Artery transverse section data sets
The training data set contains 283 images. The images were coupled with labeling data describing the positions and sizes of the arteries sought. These data were used to design the evolutionary cascade and to create training images 24×24 pixels in size. This means that 283 positive and 282 negative images were created. The training was then performed on a total of 566 images. Testing was performed on the selected data set from two ultrasound devices. Fifty video sequences from 15 adults were acquired with the Ultrasonix device; some of them were taken during inspiration or swallowing. From this sequence, every 20th frame was made part of the Ultrasonix test data set, which finally consisted of 538 test images. Twenty-two video sequences were acquired with the Toshiba device from two adults and one child. Every 10th frame was made part of the Toshiba test data set, which finally consisted of 433 test images including noisy and deformed specimens. Similarly to the training data set, all test data images were coupled with labeling information containing ground truth information—coordinates of the center and radius of the circle being sought.
If you use something from these data sets, please cite works written further:
Ing. Kamil Riha, Ph.D., firstname.lastname@example.org