| The localization and detection of anomaly in biological tissues had always been a research hotspot in the biomedical photonics' field.Accurate positioning of anomaly was a prerequisite for the application of optical imaging methods in clinical testing.At present,domestic and foreign researchers often used the multi-source and multi-detector pattern to detect anomaly in tissues.However,there were some problems in the inverse problem reconstruction,such as a long detection time,a complicated calculation process,and the influence of the sources and the detectors' distribution on the detection result.This study proposed a method for fast localization of anomaly in tissue based on differential optical density difference.The positional information of the anomaly in the differential optical density difference curve was quickly extracted by the tissue symmetry.We further studied the method of rapid reconstruction of anomaly in tissue based on deep learning.The deep learning model was used to image the anomaly in tissue,and we studied the influence of the sample distribution on the anomaly's reconstruction results.We also studied the measurable range of the deep learning model.The method effectively improved the detection speed and accuracy,and provided a reference for the clinical application of rapid non-invasive detection of anomaly in tissues.The main research contents of this thesis include:(1)We proposed a method for fast localization of anomaly in tissue based on differential optical density difference.And we verified the feasibility of differential optical density difference theory from both theoretical and experimental aspects.We compared the prediction results of anomaly with different horizontal position,depths and sizes.We proposed a method for extracting differential optical density difference features based on Gaussian fitting.The anomaly's horizontal and depth positions were determined based on the curve's peak positions.(2)The Stacked Auto-Encoder(SAE)structure was designed and applied to the rapid reconstruction of different locations and sizes' anomaly in biological tissues.The mean value of the standard deviation between the lateral predicted value and the true value of the anomaly's center point was 0.7094 mm,and the mean value of the standard deviation between the longitudinal predicted value and the true value of the anomaly's center point was 0.5721 mm.(3)We proposed a training sample data equalization distribution method to improve the image reconstruction effect of the deep learning network.The training samples were balanced from the limited data samples,which can effectively improve the model's accuracy and stability compared with the random selection method.The experimental results showed that the network obtained by the balanced distribution training samples can increase the number of samples with position prediction error less than 1 mm from 77.2%to 90.25%.(4)The reconstruction error of different positions in the phantom model was systematically analyzed to realize the measurable analysis of anomaly's reconstruction model.According to the distribution of differential optical density values in the model,the best differential optical density threshold was selected to achieve data screening of the training set.The accuracy and stability of the prediction model were further improved.The experimental process and results showed that this method can achieve fast and effective localization and reconstruction of anomaly in tissues.This method plays an important role in optical non-invasive detection of diseases such as breast cancer and skin cancer.And it provides an auxiliary tool for tissue tomography based on near-infrared spectroscopy. |
PDF Full Text Request