| Photoacoustic computed tomography(PACT)is a novel noninvasive biomedical imaging modality,which has many important and unique applications in the field of life sciences.Image reconstruction is a key step in PACT,which has a significant impact on the final image quality.To produce high-quality images,most image reconstruction algorithms require that signal acquisition in PACT should ideally meet certain conditions,such as dense-view and full-view arrangement of detector elements,and full-bandwidth detection.However,in practice,these conditions are rarely fulfilled and the system often provides incompletely sampled photoacoustic data,resulting in inaccurate image reconstruction and image quality degradation.This thesis focuses on the problem of image reconstruction from incompletely sampled data under conditions of sparse view,limited view,and limited bandwidth.The main innovative work of this thesis includes:(1)Analysis and suppression of negativity artifacts in PACT.Negative artifacts in PACT images have been noted in previous photoacoustic studies,but the formation mechanism and underlying causes of negative artifacts have not been explored.In this research,taking the filtered back projection(FBP)as the major representative,we analyse the formation mechanism of the negativity artifacts and investigate the fundamental causes of the artifacts,including sparse view,limited view,and limited bandwidth.Two image postprocessing techniques,i.e.,envelope detection and forced zeroing,are used to suppress the negativity artifacts and the characteristics of two methods are also analyzed and compared.This work can help the PACT community to better understand the characteristics of the negativity artifacts and choose an appropriate artifact-suppression method.It also provides a theoretical foundation for the development of the novel artifact-suppression strategies and artifact-free image reconstruction approaches.(2)Dual-domain enhancement model-based sparse-view image reconstruction in PACT.Here,we propose a dual-domain enhancement model(DRI-Net)for the sparseview PACT image reconstruction.The proposed DRI-Net consists of three modules,including a data-domain enhancement module(D-Net),a reconstruction module(RNet),and an image-domain enhancement module(I-Net).Different from other deeplearning methods,which use the single data-domain or image-domain information to enhance the results,the DRI-Net improves the accuracy of image reconstruction by using the data-domain and image-domain information simultaneously.In addition,the information transmission between data-domain network and image-domain network is available by designing a reconstruction module,and thus the DRI-Net can perform endto-end reconstruction.The numerical simulation and in vivo experiments demonstrate that DRI-Net can reconstruct the high-quality photoacoustic image from sparsely sampled data,where the reconstructed image is artifact-free and contains more details.The proposed DRI-Net approach largely improves the imaging quality of sparse-view PACT and is expected to help the PACT system achieve high-quality imaging with high speed and low cost.(3)Physical model-driven deep learning for sparse-view and limited-view PACT image reconstruction.In this work,we propose a physical model-driven deep-learning framework(dFBP)for high-quality PACT image reconstruction under conditions of sparse view and limited view.Compared with other deep-learning methods,dFBP method has following advantages.First,dFBP network couples the physical model of the traditional FBP algorithm,and thus has the high robustness of the traditional algorithm and the high adaptability of the deep-learning algorithm,and is interpretable.Second,another key point of dFBP network construction is that the transformation between the filtered signal in the data domain and the projected image in the image domain is realized by a two-dimensional sparse transformation matrix and a small-size three-dimensional decomposition matrix,rather than the fully connected layer commonly used in other deep-learning methods.This strategy reduces the number of network parameters by three orders of magnitude.Third,dFBP algorithm is versatile.It has no specific restrictions on the geometry of detectors,the number of array elements,and the size of reconstructed images.The results on animal and human show that dFBPbased PACT can directly reconstruct images from measured photoacoustic signals and produce high-quality images from sparse-view and limited-view data.High-quality image is an important impetus to promote the development of PACT.The work in this thesis can significantly improve the image quality of incompletely sampled data-based PACT and are of great significance to promote the further development of PACT. |
