Accurate recognition of early-stage pneumoconiosis presents significant challenges due to the irregular morphology, diffuse distribution, and small size of pulmonary lesions. Existing 2D methods struggle to focus on lesion-level 3D characteristics and inter-slice correlations in localized weak lesion regions, resulting in incomplete feature extraction and inaccurate calculation of lesion volume. To obtain complete 3D fine-grained lesion features in the entire lung, this paper proposes an early pneumoconiosis recognition network (EPRNet) to enhance fine-grained feature acquisition abilities and discover inter-slice correlations, thereby improving early pneumoconiosis recognition accuracy in a more structured and flexible manner. Specifically, to obtain the fine-grained 3D features of early pneumoconiosis more comprehensively, a distance-similarity graph encoding module is proposed to construct and encode the relationships of the distributed tiny lesions within CT slices, integrating the spatial positions and the corresponding feature similarities of the lesions to improve the accuracy of pneumoconiosis feature representation. To adaptively preserve accurate graph representations of the correlations of the lesions across inter-CT slices, a hierarchical dynamic-scored adaptive pooling module is proposed to discover the potential long distance correlations between cross-slices, obtaining spatial semantic information of diffused lesions in the entire lung. Experimental results based on multiple datasets demonstrate that EPRNet achieves state-of-the-art performance while exhibiting better generalization. The ablation experiments also prove the effectiveness of each module in EPRNet.

Ultrasound imaging and biochemical examinations are the primary methods for diagnosing Hashimoto's thyroiditis (HT). However, neither of them is sufficient to accurately diagnose HT alone. Most existing multimodal models for HT diagnosis focus primarily on extracting and concatenating features from different modalities, which are ineffective due to the dimensional imbalance of the features between the textual and image data. To address this issue, we propose a novel Multimodal Collaborative Fusion Learning (MCFL) approach, which can enhance and recalibrate the biochemical indicators using ultrasound images, effectively improving the significance and specificity of biochemical indicators for the diagnosis of HT. Specifically, MCFL first constructs a novel INNet to convert the image-level characteristics of the HT ultrasound image into two numerical indicators, i.e., the Local prominent inflammatory (Lpi) and the Global diffuse lesion (Gdl), unifying image data and textual data into a single representation space. Then, a decision tree-based optimization strategy is employed to supervise the training of INNet, interactively recalibrating biochemical indicators with the guidance of the two numerical indicators mentioned above and obtaining a more accurate feature representation of HT. Finally, based on the deep Q-learning framework, a reward mechanism is established to guide the HT diagnostic process, in which the experience replay mechanism and the $\epsilon $ -greedy strategy are utilized collaboratively to improve the accuracy and robustness of the model. Extensive experiments are conducted on a multimodal dataset from multiple medical centers, and the results demonstrate that MCFL achieves state-of-the-art performance, setting a new benchmark.

Metal implants introduce severe artifacts in CT images, compromising diagnostic reliability. Supervised metal artifact reduction (MAR) models trained on simulated data are effective but often fail due to domain gaps when applied to real clinical data. Unsupervised methods trained on real images avoid such gaps but suffer from weak artifact suppression and training instability. To address these challenges, we propose d-MAR, a novel MAR framework that performs diffusion-driven domain transformations between simulated and real image domains. Specifically, real image domain (RID) data is transformed into the simulated image domain (SID), processed by a MAR model trained on simulation-paired data, and transformed back into RID. We harness diffusion models as a transformation bridge and introduce two targeted conditional sampling techniques-conditional input and sampling enhancement-based on Fourier-extracted low-frequency image components. This enables domain alignment without random generation, ensuring consistent anatomical fidelity. The proposed d-MAR can reduce real metal artifacts originating from different scanning protocols and devices with a MAR model trained with simulated paired data. Evaluations on Clinical Head, Clinical Body, and dental CBCT datasets show that d-MAR consistently outperforms conventional MAR methods in both quantitative metrics and visual quality, demonstrating strong generalization capability.

Although supervised deep learning methods have made significant advances in low-dose computed tomography (LDCT) image denoising, these approaches typically require pairs of low-dose and normal-dose CT images for training, which are often unavailable in clinical settings. Self-supervised deep learning (SSDL) has great potential to cast off the dependence on paired training datasets. However, existing SSDL methods are limited by the neighboring noise independence assumptions, making them ineffective for handling spatially correlated noises in LDCT images. To address this issue, this paper introduces a novel SSDL approach, named, Noise-Aware Blind Spot Network (NA-BSN), for high-quality LDCT imaging, while mitigating the dependence on the assumption of neighboring noise independence. NA-BSN achieves high-quality image reconstruction without referencing clean data through its explicit noise-aware constraint mechanism during the self-supervised learning process. Specifically, it is experimentally observed and theoretical proven that the $l1$ norm value of CT images in a downsampled space follows a certain descend trend with increasing of the radiation dose, which is then used to construct the explicit noise-aware constraint in the architecture of BSN for self-supervised LDCT image denoising. Various clinical datasets are adopted to validate the performance of the presented NA-BSN method. Experimental results reveal that NA-BSN significantly reduces the spatially correlated CT noises and retains crucial image details in various complex scenarios, such as different types of scanning machines, scanning positions, dose-level settings, and reconstruction kernels.

Ultrasound imaging can reveal the typical changes in thyroid tissue caused by Hashimoto's Thyroiditis (HT), which plays a crucial role in HT diagnosis. Clinicians should perform ultrasound imaging from multiple anatomical planes to obtain a series of thyroid images, thereby collecting comprehensive information about HT lesions from different perspectives. However, ultrasound images lacking typical HT characteristics or containing extensive non-HT regions may hinder the identification of localized lesions, thereby increasing the complexity of diagnosis. In addition, the ability to capture synergistic interactions among multiple lesions in critical ultrasound images is crucial for improving the diagnostic accuracy of HT. To address these challenges, a novel weakly supervised multi-instance learning model, HTMIL, is proposed for HT diagnosis, requiring only patient-level data annotation. HTMIL consists of a Global-Neighbor extraction Layer (GNL) and a Cross-Aggregation Layer (CAL). Specifically, a Global-Neighbor Awareness (GNA) module is proposed in GNL to allow HTMIL to focus on key localized lesions and ignore noise from unrelated regions in HT ultrasound images, thus improving the effectiveness of extracting focused features from localized lesions. The Adaptive Bilinear Pooling (ABP) module is introduced in CAL to capture the interaction features between key localized lesions and other characteristics of HT in critical ultrasound images, thus achieving synergistic diagnosis and further increasing diagnostic accuracy. HTMIL achieves state-of-the-art (SOTA) performance with 87.67% accuracy and 91.73% AUC on a multicenter HT dataset, and its robustness is further validated on a public dataset.

Distinguishing Hashimoto's thyroiditis (HT) lesions from ordinary thyroid tissues is difficult with ultrasound images. Challenges in achieving high performance of HT ultrasound image classification include the low resolution, blurred features and large area of irrelevant noise. To address these problems, we propose a Feature-level Boosting Ensemble Network (FBENet) for HT ultrasound image classification. Specifically, to capture the features of suspicious HT lesions efficiently, an Ensemble Feature Boosting Module (EFBM) is introduced into the feature-level ensemble to boost the blurred features. Then, the spatial attention mechanism is adopted in backbone models to improve the feature focusing performance and representation ability. Furthermore, feature-level ensemble technique is employed in the training process to achieve more comprehensive feature representation ability. Experimentally, FBENet was trained on 6,503 HT ultrasound images, and tested on 1,626 HT ultrasound images with 82.92% accuracy and 89.24% AUC on average.

Supervised deep-learning techniques with paired training datasets have been widely studied for low-dose computed tomography (LDCT) imaging with excellent performance. However, the paired training datasets are usually difficult to obtain in clinical routine, which restricts the wide adoption of supervised deep-learning techniques in clinical practices. To address this issue, a general idea is to construct a pseudo paired training dataset based on the widely available unpaired data, after which, supervised deep-learning techniques can be adopted for improving the LDCT imaging performance by training on the pseudo paired training dataset. However, due to the complexity of noise properties in CT imaging, the LDCT data are difficult to generate in order to construct the pseudo paired training dataset. In this article, we propose a simple yet effective cross-domain unpaired learning framework for pseudo LDCT data generation and LDCT image reconstruction, which is denoted as CrossDuL. Specifically, a dedicated pseudo LDCT sinogram generative module is constructed based on a data-dependent noise model in the sinogram domain, and then instead of in the sinogram domain, a pseudo paired dataset is constructed in the image domain to train an LDCT image restoration module. To validate the effectiveness of the proposed framework, clinical datasets are adopted. Experimental results demonstrate that the CrossDuL framework can obtain promising LDCT imaging performance in both quantitative and qualitative measurements.

Deep learning (DL)-based methods show great potential in computed tomography (CT) imaging field. The DL-based reconstruction methods are usually evaluated on the training and testing datasets which are obtained from the same distribution, i.e., the same CT scan protocol (i.e., the region setting, kVp, mAs, etc.). In this work, we focus on analyzing the robustness of the DL-based methods against protocol-specific distribution shifts (i.e., the training and testing datasets are from different region settings, different kVp settings, or different mAs settings, respectively). The results show that the DL-based reconstruction methods are sensitive to the protocol-specific perturbations which can be attributed to the noise distribution shift between the training and testing datasets. Based on these findings, we presented a low-dose CT reconstruction method using an unsupervised strategy with the consideration of noise distribution to address the issue of protocol-specific perturbations. Specifically, unpaired sinogram data is enrolled into the network training, which represents unique information for specific imaging protocol, and a Gaussian mixture model (GMM) is introduced to characterize the noise distribution in CT images. It can be termed as GMM based unsupervised CT reconstruction network (GMM-unNet) method. Moreover, an expectation-maximization algorithm is designed to optimize the presented GMM-unNet method. Extensive experiments are performed on three datasets from different scan protocols, which demonstrate that the presented GMM-unNet method outperforms the competing methods both qualitatively and quantitatively.

X-ray computed tomography (CT) is widely used clinically to diagnose a variety of diseases by reconstructing the tomographic images of a living subject using penetrating X-rays. For accurate CT image reconstruction, a precise imaging geometric model for the radiation attenuation process is usually required to solve the inversion problem of CT scanning, which encodes the subject into a set of intermediate representations in different angular positions. Here, we show that accurate CT image reconstruction can be subsequently achieved by downsampled imaging geometric modeling via deep-learning techniques. Specifically, we first propose a downsampled imaging geometric modeling approach for the data acquisition process and then incorporate it into a hierarchical neural network, which simultaneously combines both geometric modeling knowledge of the CT imaging system and prior knowledge gained from a data-driven training process for accurate CT image reconstruction. The proposed neural network is denoted as DSigNet, i.e., downsampled-imaging-geometry-based network for CT image reconstruction. We demonstrate the feasibility of the proposed DSigNet for accurate CT image reconstruction with clinical patient data. In addition to improving the CT image quality, the proposed DSigNet might help reduce the computational complexity and accelerate the reconstruction speed for modern CT imaging systems.

The Radon transform is widely used in physical and life sciences, and one of its major applications is in medical X-ray computed tomography (CT), which is significantly important in disease screening and diagnosis. In this paper, we propose a novel reconstruction framework for Radon inversion with deep learning (DL) techniques. For simplicity, the proposed framework is denoted as iRadonMAP, i.e., inverse Radon transform approximation. Specifically, we construct an interpretable neural network that contains three dedicated components. The first component is a fully connected filtering (FCF) layer along the rotation angle direction in the sinogram domain, and the second one is a sinusoidal back-projection (SBP) layer, which back-projects the filtered sinogram data into the spatial domain. Next, a common network structure is added to further improve the overall performance. iRadonMAP is first pretrained on a large number of generic images from the ImageNet database and then fine-tuned with clinical patient data. The experimental results demonstrate the feasibility of the proposed iRadonMAP framework for Radon inversion.

Reducing the exposure to X-ray radiation while maintaining a clinically acceptable image quality is desirable in various CT applications. To realize low-dose CT (LdCT) imaging, model-based iterative reconstruction (MBIR) algorithms are widely adopted, but they require proper prior knowledge assumptions in the sinogram and/or image domains and involve tedious manual optimization of multiple parameters. In this paper, we propose a deep learning (DL)-based strategy for MBIR to simultaneously address prior knowledge design and MBIR parameter selection in one optimization framework. Specifically, a parameterized plug-and-play alternating direction method of multipliers (3pADMM) is proposed for the general penalized weighted least-squares model, and then, by adopting the basic idea of DL, the parameterized plug-and-play (3p) prior and the related parameters are optimized simultaneously in a single framework using a large number of training data. The main contribution of this paper is that the 3p prior and the related parameters in the proposed 3pADMM framework can be supervised and optimized simultaneously to achieve robust LdCT reconstruction performance. Experimental results obtained on clinical patient datasets demonstrate that the proposed method can achieve promising gains over existing algorithms for LdCT image reconstruction in terms of noise-induced artifact suppression and edge detail preservation.