Communication Sciences

Ol?jecti~'e: Development of an artificial intelligent diagnostic system for the interpretation of umbilical artery blood flow velocity waveform measurements. Study design: Study design comprised several stages including data acquisition, image processing and analysis, training of artificial neural network and testing the predictive value of the system. The clinical material was handled in two groups. The training group consisted of 952 umbilical artery blood flow velocity waveform images of 174 normal pregnancies with normal outcome, while the testing group was composed of 138 images derived from 20 normal pregnancies with normal outcome and 68 images of 16 high risk pregnancies with poor outcome. All subjects were evaluated by Doppler ultrasonography and umbilical artery blood flow velocity waveform images were transferred to the computer environment by means of a special data acquisition system. Automated image processing and analysis were performed to derive indices such as A/B ratio, resistance index, pulse related index, area ratio of wave and angle of coincident slopes. We have used a supervised artificial neural network (back propagation learning algorithm) to develop an intelligent diagnostic system which is called the BOLU system. Results: This version of the system was trained with the umbilical artery blood flow velocity waveform images of normal pregnancies. Thus, the BOLU system decides whether the tested image is normal for a given gestational week or not. The specificity and sensitivity of this system were estimated to be 98.6% and 51.5% respectively. Conclusion: We have developed an artificial intelligent diagnostic system for the interpretation of umbilical artery blood flow velocity waveform measurements. Waveform indices were obtained automatically by image processing and analysis. The predictive value of the system was found to be satisfactory.

Wavelet transformations of images have the valuable properties of providing robust data compression and image details that can be used for discerning signal features at different resolutions. With the approach presented here, such features, represented by wavelet coefficients, are extracted from the variations of blood velocity waveforms obtained from Doppler ultrasound images of fetal umbilical arteries. The reliability of these features are tested and proven by image reconstruction from wavelet coefficients thresholded according to the required resolution. The obtained reliable features form the training samples of an artificial neural network to be used in intelligent diagnostics for fetal surveillance

We introduce a hybrid intelligent karyotyping system based on two different types of artificial neural networks (ANNs) and chromosome's features obtained by digital image processing techniques. A microscope equipped with a CCTV camera and a microcomputer including a frame grabber are the basic components of our hardware set-up. The inputs to the ANN structure are obtained directly from digital chromosome images by using two recently developed object detection and object skeletonizing algorithms. Moreover, the band patterns of chromosomes are represented by applying wavelet transform techniques on the gray level profiles of chromosomes. The network parameters are determined by using the results of many training and testing experiments in order to reach an optimal state from the classification performance point of view