The Research For Early Detection And Degree Of Liver Fibrosis Recognization Based On Ultrasound Radio Frequency Time Series

Early detection and degree recognization of liver fibrosis have clinical value for preventing cirrhosis of liver and dynamically monitoring the progress of liver diseases. Even though tissue biopsy serves as the gold standard for determining the extent of liver fibrosis, it is still not accepted by patients, for it is invasive, and its accuracy is subject to uniformity of fibrosis. Thus, a clinically alternative non-invasive and effective method to assess the level of liver fibrosis is sorely needed. Therefore, a recognition model for early liver fibrosis is proposed in this paper, using ultrasonic tissue characterization and data mining strategies, based on ultrasound RF time series extracted from vivo rat liver samples at different pathological degrees. The main propose of this paper is to inverstigate the possibility of early detecting liver fibrosis and predicting its degree with a non-invasive means. Detailed work of the thesis includes:1) Due to the unknown noise type of ultrasound RF time series, an improved adaptive algorithm to determine the optimal decomposition level in wavelet transform is proposed2) The effects of noise on signal distribution and the precision of ultrasonic tissue characterization were analyzed. Four situations of signal pre-treatment are setted, respectively unprocessed, wavelet denoising, spatial denoising, wavelet denoising after spatial denoising. In each situation, distribution type of ultrasound RF time series is tested, and features are extracted, after which the degree of liver fibrosis is identified. The results show that wavelet denoising has little impact on signal distribution, whereas spatial denoising has larger influence. After denoising, the model precision is becoming lower, indicating that noise carrys “tissue characterizing” information.3) For the first time, the data distribution of ultrasound RF time series is reported. Based on Kolmogorov-Smirnov(KS) test, an experiment was designed to test the distribution of ultrasound RF time series. In this experiment, 9 assumptions of distribution type were applied. The result is as follows: In situations of unprocessed and wavelet denoising, chi-square distribution is the distrubution of ultrasound RF time series.4) 10 methods of computing fractal dimension were compared by simulation experiment. Calculation error, signal length and signal type were used to estimate each method. After that, Higuchi method is selected to calculate fractal dimension of ultrasound RF time series, and the optimum length of ultrasound RF time series is determined.5) Fractal dimension, 6 frequency features and 5 time features were extracted from ultrasound RF time series. The Analysis of performance of these features by box-plot shows that a single feature can not distinguish liver fibrosis degree between 0~4 effectively.6) Three algorithms about feature fusion were proposed. Iris data set, which is a standard database in the machine learning literature, SVM and Random Forest classifiers were used to verify the usefulness of these algorithms. It turns out that these proposed algorithms have higher accuracy with Random Forest classifier.7) Based on SVM and Random Forest classifiers, recognition models for early liver fibrosis were built in 24 cases. The highest classification accuracy reached 96.6667%.8) A recognition system for Liver Fibrosis was developed and has been applicated in Sun Yat Sen University Cancer Center. It adopted such techniques as MFC, MySQL and OpenCV.