Machine learning aided bioimpedance tomography for tissue engineering

In tissue engineering, miniature Electrical Impedance Tomography (mEIT) (or bioimpedance tomography), is an emerging tomographic modality that contributes to non-destructive and label-free imaging and monitoring of 3-D cellular dynamics. The main challenge of mEIT comes from the nonlinear and ill-posed image reconstruction problem, leading to the increased sensitivity to imperfect measurement signals. Physical model-based image reconstruction methods have been successfully applied to conventional setups, but are less satisfying for the mEIT setup regarding image quality, conductivity retrieval and computational efficiency. Data-driven or learning-based methods have recently become a new frontier for tomographic image reconstruction, particularly for medical imaging modalities, e.g., Computed Tomography (CT), Positron Emission Tomography (PET), and Magnetic Resonance Imaging (MRI). However, the study of learning-based image reconstruction methods in challenging micro-scale sensor setups and the seamless integration of such algorithms with the tomography instrument remains a gap. This thesis aims to develop 2-D and 3-D imaging platforms integrating multi-frequency EIT and machine learning-based image reconstruction algorithms to extract spectroscopic electrical properties of 3-D cultivated cells under in vitro conditions, in a non-destructive, robust, and computation-efficient manner. Recent advances in deep learning have pointed out a promising alternative for EIT image reconstruction. However, it is still challenging to image multiple objects with varying conductivity levels with a single neural network. A deep learning and group sparsity regularization-based hybrid image reconstruction framework was proposed to enable high-quality cell culture imaging with mEIT. A deep neural network was proposed to estimate the structural information in binary masks, given the limited number of data sets. Then the structural information is encoded in group sparsity regularization to obtain the final conductivity estimation. We validated our approach by imaging 3D cancer cell spheroids (MCF-7). Our method can be readily translated to spheroids, organoids, and cell culture in scaffolds of biomaterials. As the measured conductivity is a proxy for cell viability, mEIT has excellent potential to enable non-invasive, real-time, long-term monitoring of 3D cell growth, opening new avenues in regenerative medicine and drug testing. Deep learning provides binary structural information in the above-mentioned hybrid learning approach, whereas the regularization algorithm determines conductivity contrasts. Despite the advancement of structure distribution, the exact conductivity values of different objects are less accurately estimated by the regularization-based framework, which essentially prevents EIT's transition from generating qualitative images to quantitative images. A structure-aware dual-branch deep learning method was proposed to further tackle this issue to predict structure distribution and conductivity values. The proposed network comprises two independent branches to encode the structure and conductivity features, respectively, and the two branches are joined later to make final predictions of conductivity distributions. Numerical and experimental evaluation results on MCF-7 human breast cancer cell spheroids demonstrate the superior performance of the proposed method in dealing with the multi-level, continuous conductivity reconstruction problem. Multi-frequency Electrical Impedance Tomography (mfEIT) is an emerging biomedical imaging modality to reveal frequency-dependent conductivity distributions in biomedical applications. Conventional model-based image reconstruction methods suffer from low spatial resolution, unconstrained frequency correlation and high computational cost. Most existing learning-based approaches deal with the single-frequency setup, which is inefficient and ineffective when extended to the multi-frequency setup. A Multiple Measurement Vector (MMV) model-based learning algorithm named MMV-Net was proposed to solve the mfEIT image reconstruction problem. MMV-Net considers the correlations between mfEIT images and unfolds the update steps of the Alternating Direction Method of Multipliers for the MMV problem (MMV-ADMM). The nonlinear shrinkage operator associated with the weighted l_{1,2} regularization term of MMV-ADMM is generalized in MMV-Net with a cascade of a Spatial Self-Attention module and a Convolutional Long Short-Term Memory (ConvLSTM) module to capture intra- and inter-frequency dependencies better. The proposed MMV-Net was validated on our Edinburgh mfEIT Dataset and a series of comprehensive experiments. The results show superior image quality, convergence performance, noise robustness and computational efficiency against the conventional MMV-ADMM and the state-of-the-art deep learning methods. Finally, few work on image reconstruction for Electrical Impedance Tomography (EIT) focuses on 3D geometries. Existing reconstruction algorithms adopt voxel grids for representation, which typically results in low image quality and considerable computational cost, and limits their applicability to real-time applications. In contrast, point clouds are a more efficient format for 3D surfaces, and such representation can naturally handle 3D shapes of arbitrary topologies with fine-grained details. Therefore, a learning-based 3D EIT reconstruction algorithm with efficient 3D representations (i.e., point cloud) was proposed to achieve higher image accuracy, spatial resolution and computational efficiency. A transformer-like point cloud network is adopted for 3D image reconstruction. This network simultaneously recovers the 3D coordinates of points to adaptively portray the objects' surface and predicts each point's conductivity. The results show that point cloud provides more efficient fine-shape descriptions and effectively alleviates computational costs. In summary, the work demonstrated in this thesis addressed the research void in tissue imaging with bioimpedance tomography by developing learning-based imaging approaches. The results achieved in this thesis could promote bioimpedance tomography as a robust, intelligent imaging technique for tissue engineering applications.