Deep learning technology has been utilized in computed tomography, but, it needs centralized dataset to train the neural networks. To solve it, federated learning has been proposed, which collaborates the data from different local medical institutions with privacy-preserving decentralized strategy. However, lots of unpaired data is not included in the local models training and directly aggregating the parameters would degrade the performance of the updated global model. In order to deal with the issues, we present a semi-supervised and semi-centralized federated learning method to promote the performance of the learned global model. Specifically, each local model is trained with an unsupervised strategy locally at a fixed round. After that, the parameters of the local models are shared to aggregate on the server to update the global model. Then, the global model is further trained with a standard dataset, which contains well paired training samples to stabilize and standardize the global model. Finally, the global model is distributed to local models for the next training step. For shorten, we call the presented federated learning method as “3SC-FL”. Experiments demonstrate the presented 3SC-FL outperforms the compared methods, qualitatively and quantitatively.

Objective. Metal artifacts in the computed tomography (CT) imaging are unavoidably adverse to the clinical diagnosis and treatment outcomes. Most metal artifact reduction (MAR) methods easily result in the over-smoothing problem and loss of structure details near the metal implants, especially for these metal implants with irregular elongated shapes. To address this problem, we present the physics-informed sinogram completion (PISC) method for MAR in CT imaging, to reduce metal artifacts and recover more structural textures. Approach. Specifically, the original uncorrected sinogram is firstly completed by the normalized linear interpolation algorithm to reduce metal artifacts. Simultaneously, the uncorrected sinogram is also corrected based on the beam-hardening correction physical model, to recover the latent structure information in metal trajectory region by leveraging the attenuation characteristics of different materials. Both corrected sinograms are fused with the pixel-wise adaptive weights, which are manually designed according to the shape and material information of metal implants. To furtherly reduce artifacts and improve the CT image quality, a post-processing frequency split algorithm is adopted to yield the final corrected CT image after reconstructing the fused sinogram. Main results. We qualitatively and quantitatively evaluated the presented PISC method on two simulated datasets and three real datasets. All results demonstrate that the presented PISC method can effectively correct the metal implants with various shapes and materials, in terms of artifact suppression and structure preservation. Significance. We proposed a sinogram-domain MAR method to compensate for the over smoothing problem existing in most MAR methods by taking advantage of the physical prior knowledge, which has the potential to improve the performance of the deep learning based MAR approaches.

Federated learning method shows great potential in computed tomography imaging field by utilizing a decentralized strategy with data privacy-preserving for local medical institutions. However, directly aggregating the parameters of each local model would degrade the generalization performance of the updated global model. In addition, well paired centralized training datasets can be collected in real world, which are not included in the current federated learning methods. To address the issue, we present a semi-centralized federated learning method to promote the generalization performance of the learned global model. Specifically, each local model is firstly trained locally at a fixed round, then, the parameters are aggregated on server to initialized the global model. After that, the global model is further trained with a standard dataset on the server, which contains well paired training samples to stabilize and standardize the global model. For shorten, we call the presented semi-centralized federated learning method as “SC-FL”. Experimental results on different local datasets demonstrate the presented SC-FL outperforms the competing methods.

Computed tomography (CT) is a widely used medical imaging modality which is capable of displaying the fine details of human body. In clinics, the CT images need to highlight different desired details or structures with different filter kernels and different display windows. To achieve this goal, in this work, we proposed a deep learning based ”All-in-One” (DAIO) combined visualization strategy for high-performance disease screening in the disease screening task. Specifically, the presented DAIO method takes into consideration of both kernel conversion and display window mapping in the deep learning network. First, the sharp kernel, smooth kernel reconstructed images and lung mask are collected for network training. Then, the structure is adaptively transferred to the kernel style through local kernel conversion to make the image have higher diagnostic value. Finally, the dynamic range of the image is compressed to a limited gray level by the mapping operator based on the traditional window settings. Moreover, to promote the structure details enhancement, we introduce a weighted mean filtering loss function. In the experiment, nine of the ten full dose patients cases from the Mayo clinic dataset are utilized to train the presented DAIO method, and one patient case from the Mayo clinic dataset are used for test. Results shows that the proposed DAIO method can merge multiple kernels and multiple window settings into a single one for the disease screening.

Dual Energy CT (DECT) has ability to characterize different materials and quantify the densities or proportions of different contrast agents. However, the basis images decomposition is an ill-posed problem and the traditional model-based and image-domain direct inversion methods always suffer from serious degradation of the signal-to-noise ratios (SNRs). To this issue, we propose a new strategy by combining model-based and learning-based methods, which suppresses noise in the material images after direct inverse, and design a semi-supervised framework, Adaptive Semi-supervised Learning Material Estimation Network (ASLME-Net), to balance the detail structure preservation and noise suppression when fed little paired data in training stage of the deep learning. Specifically, the ASLME-Net contains two sub-networks, i.e., supervise sub-network and unsupervised sub-network. The supervised sub-network aims at capturing key features learned by with the labeled data, and the unsupervised sub-network adaptively learns the transferred feature distribution from supervised sub-network with Kullback-Leibler (KL) divergence. Experiment shows that the presented method can suppress the noise propagation in decomposition and yield qualitatively and quantitatively accurate results during the process of material decomposition. To this issue, we propose a new strategy by combining model-based and learning-based methods, which suppresses noise in the material images after direct inverse, and design a semi-supervised framework, Adaptive Semi-supervised Learning Material Estimation Network (ASLME-Net), to balance the detail structure preservation and noise suppression when fed little paired data in training stage of the deep learning.