The multi-classification of histopathological images under imbalanced sample conditions remains a long-standing unresolved challenge in computational pathology. In this paper, we propose for the first time a cross-patient pseudo-bag generation technique to address this challenge. Our key innovation lies in a cross-patient pseudo-bag generation framework that extracts complementary pathological features to construct distributionally consistent pseudo-bags. To resolve the critical challenge of distributional alignment in pseudo-bag generation, we propose an affinity-driven curriculum contrastive learning strategy, integrating sample affinity metrics with progressive training to stabilize representation learning. Unlike prior methods focused on bag-level embeddings, our framework pioneers a paradigm shift toward multi-instance feature distribution mining, explicitly modeling inter-bag heterogeneity to address class imbalance. Our method demonstrates significant performance improvements on three datasets with multiple classification difficulties, outperforming the second-best method by an average of 1.95 percentage points in F1 score and 2.07 percentage points in ACC.

In recent years, dynamic functional connectivity (dFC) has been widely employed for brain disease diagnosis. By leveraging the inherent topological characteristics of the brain, graph neural networks (GNNs) have emerged as prominent deep learning methods for utilizing dFC in this context. However, existing research has some limitations. Temporally, the conventional fixed-length sliding window approach often fails to capture the multi-scale temporal characteristics inherent in brain activity. Spatially, GNN-derived graph representations usually overlook the multi-network participation of brain regions. To address these limitations, we propose Ada-MST, an adaptive multi-scale spatio-temporal model utilizing multi-scale dFC for brain disease diagnosis. Our framework constructs personalized multi-scale dFC graphs that adapt to subject-specific temporal characteristics. Moreover, we introduce a novel overlapping community-aware readout module that incorporates the participation of brain regions in multiple functional networks, leading to more accurate graph-level representations. Experiments on ABIDE-I and ABIDE-II datasets demonstrate that our method outperforms state-of-the-art approaches. Visualization analysis further confirms the generalizability of the subject-adaptive graphs and their focus on disease-related brain activity. Furthermore, the fuzzy memberships revealed by our readout module indicate distinct patterns across diseases, suggesting the promise of considering functional community membership changes for exploring disease biomarkers.

Segmentation of complex medical images such as vascular network and pulmonary tracheal network requires segmentation of many tiny targets on each tomographic section of the 3-D medical image volume. Although semantic segmentation of medical images based on deep learning has made great progress, fully supervised models require a great amount of annotations, making such complex medical image segmentation a difficult problem. In this article, we propose a semi-supervised model for complex medical image segmentation, which innovatively proposes a bidirectional self-training paradigm, through dynamically exchanging the roles of teacher and student by estimating the reliability at the model level. The direction of information and knowledge transfer between the two networks can be controlled, and the probability distribution of the roles of teacher and student in the next stage will be jointly determined by the model's uncertainty and instability in the training process. We also resolve the problem that loosely coupled networks are prone to collapse when training on small-scale annotated data by proposing asymmetric supervision (AS) strategy and hierarchical dual student (HDS) structure. In particular, a bidirectional distillation loss combined with the role exchange (RE) strategy and a global-local-aware consistency loss are introduced to obtain stable mutual promotion and achieve matching of global and local features, respectively. We conduct detailed experiments on two public datasets and one private dataset and lead existing semi-supervised methods by a large margin, while achieving fully supervised performance at a labeling cost of 5%.

Spinal X-ray effectively visualizes the overall spinal situation and vertebral details. However, X-ray is unsuitable for long-term follow-up or frequent monitoring due to its radiation hazard. Motivated by this, we propose a model named dual-decomposition radiograph synthesizer (DDRS) to predict the X-ray image of the present moment, given the previous X-ray image and a pair of three-dimensional spinal ultrasound data, for a practical and radiation-free evaluation of spinal deformity in follow-up or monitoring. The DDRS used a novel dual-decomposition strategy to ensure the quality of synthesized images. First, the DDRS innovatively converted the X-ray image synthesis into a fusion between the spinal pose and anatomical information. A parallel architecture was used to extract and aggregate the two information from ultrasound and X-ray images. Second, an intermodality calibration module and a global-local cooperated feature extractor are further introduced to implement our synthesis strategy effectively. The intermodality calibration module provides an accurate spinal pose description by correcting a potential pose difference during two image acquisition times. The global-local cooperated feature extractor contributes to preserving spinal anatomical information in the previous X-ray image by exploring global dependencies and highlighting local details. Extensive experiments were conducted on a real clinical dataset. Results show that a mean structural similarity (SSIM) of 0.89 was obtained between synthesized X-ray images provided by the DDRS and real ones, and further comparisons with existing outstanding image synthesizers also display a 17.1% improvement in mean SSIM, illustrating the potential of our synthesizer in a radiation-free follow-up of scoliosis.

BACKGROUND: Brain tumor segmentation is a challenging problem in medical image processing and analysis. It is a very time-consuming and error-prone task. In order to reduce the burden on physicians and improve the segmentation accuracy, the computer-aided detection (CAD) systems need to be developed. Due to the powerful feature learning ability of the deep learning technology, many deep learning-based methods have been applied to the brain tumor segmentation CAD systems and achieved satisfactory accuracy. However, deep learning neural networks have high computational complexity, and the brain tumor segmentation process consumes significant time. Therefore, in order to achieve the high segmentation accuracy of brain tumors and obtain the segmentation results efficiently, it is very demanding to speed up the segmentation process of brain tumors. RESULTS: Compared with traditional computing platforms, the proposed FPGA accelerator has greatly improved the speed and the power consumption. Based on the BraTS19 and BraTS20 dataset, our FPGA-based brain tumor segmentation accelerator is 5.21 and 44.47 times faster than the TITAN V GPU and the Xeon CPU. In addition, by comparing energy efficiency, our design can achieve 11.22 and 82.33 times energy efficiency than GPU and CPU, respectively. CONCLUSION: We quantize and retrain the neural network for brain tumor segmentation and merge batch normalization layers to reduce the parameter size and computational complexity. The FPGA-based brain tumor segmentation accelerator is designed to map the quantized neural network model. The accelerator can increase the segmentation speed and reduce the power consumption on the basis of ensuring high accuracy which provides a new direction for the automatic segmentation and remote diagnosis of brain tumors.

Thin-section magnetic resonance imaging (MRI) can provide higher resolution anatomical structures and more precise clinical information than thick-section images. However, thin-section MRI is not always available due to the imaging cost issue. In multicenter retrospective studies, a large number of data are often in thick-section manner with different section thickness. The lack of thin-section data and the difference in section thickness bring considerable difficulties in the study based on the image big data. In this article, we introduce DeepVolume, a two-step deep learning architecture to address the challenge of accurate thin-section MR image reconstruction. The first stage is the brain structure-aware network, in which the thick-section MR images in axial and sagittal planes are fused by a multitask 3-D U-net with prior knowledge of brain volume segmentation, which encourages the reconstruction result to have correct brain structure. The second stage is the spatial connection-aware network, in which the preliminary reconstruction results are adjusted slice-by-slice by a recurrent convolutional network embedding convolutional long short-term memory (LSTM) block, which enhances the precision of the reconstruction by utilizing the previously unassessed sagittal information. We used 305 paired brain MRI samples with thickness of 1.0 mm and 6.5 mm in this article. Extensive experiments illustrate that DeepVolume can produce the state-of-the-art reconstruction results by embedding more anatomical knowledge. Furthermore, considering DeepVolume as an intermediate step, the practical and clinical value of our method is validated by applying the brain volume estimation and voxel-based morphometry. The results show that DeepVolume can provide much more reliable brain volume estimation in the normalized space based on the thick-section MR images compared with the traditional solutions.

In magnetic resonance imaging (MRI), different imaging settings lead to various intensity distributions for a specific imaging object, which brings huge diversity to data-driven medical applications. To standardize the intensity distribution of magnetic resonance (MR) images from multiple centers and multiple machines using one model, a cycle generative adversarial network (CycleGAN)-based framework is proposed. It utilizes a unified forward generative adversarial network (GAN) path and multiple independent backward GAN paths to transform images in different groups into a single reference one. To preserve image details and prevent resolution loss, two jump connections are applied in the CycleGAN generators. A weak-pair strategy is designed to fully utilize the prior knowledge of the organ structure and promote the performance of the GANs. The experiments were conducted on a T2-FLAIR image database with 8192 slices from 489 patients. The database was obtained from four hospitals and five MRI scanners and was divided into nine groups with different imaging parameters. Compared with the representative algorithms, the peak signal-to-noise ratio, the histogram correlation, and the structural similarity were increased by 3.7%, 5.1%, and 0.1% on average, respectively; the gradient magnitude similarity deviation, the mean square error, and the average disparity were reduced by 19.0%, 15.7%, and 9.9% on average, respectively. Experiments also showed the robustness of the proposed model with a different training set configuration and effectiveness of the proposed framework over the original CycleGAN. Therefore, the MR images with different imaging settings could be efficiently standardized by the proposed method, which would benefit various data-driven applications.

Brain tumors are the most common malignant neurologic tumors with the highest mortality and disability rate. Because of the delicate structure of the brain, the clinical use of several commonly used biopsy diagnosis is limited for brain tumors. Radiomics is an emerging technique for noninvasive diagnosis based on quantitative medical image analyses. However, current radiomics techniques are not standardized regarding feature extraction, feature selection, and decision making. In this paper, we propose a sparse representation-based radiomics (SRR) system for the diagnosis of brain tumors. First, we developed a dictionary learning- and sparse representation-based feature extraction method that exploits the statistical characteristics of the lesion area, leading to fine and more effective feature extraction compared with the traditional explicitly calculation-based methods. Then, we set up an iterative sparse representation method to solve the redundancy problem of the extracted features. Finally, we proposed a novel multi-feature collaborative sparse representation classification framework that introduces a new coefficient of regularization term to combine features from multi-modal images at the sparse representation coefficient level. Two clinical problems were used to validate the performance and usefulness of the proposed SRR system. One was the differential diagnosis between primary central nervous system lymphoma (PCNSL) and glioblastoma (GBM), and the other was isocitrate dehydrogenase 1 estimation for gliomas. The SRR system had superior PCNSL and GBM differentiation performance compared with some advanced imaging techniques and yielded 11% better performance for estimating IDH1 compared with the traditional radiomics methods.

Ultrasonography has been widely used to evaluate duodenogastric reflux (DGR). But to the best of our knowledge, no automatic analysis system was developed to realize the quantitative computer-aided analysis. In this paper, we propose a system to perform the automatic detection of DGR in the ultrasonic image sequences by applying the automatic motion analysis. The motion field is estimated based on image velocimetry. Then, an intelligent motion analysis is applied. For the DGR detection, the motion and structural information is combined to analyze the transploric motion of the fluid. In order to test the performance of the proposed system, we designed the experiment with the real and synthetic ultrasonic data. The proposed system achieved a good performance in the DGR detection. The automatic results were accordant with the gold standard in analyzing the fluid motion. The proposed system is supposed to be a promising tool for the study and evaluation of DGR.

Tracking pylorus in ultrasonic image sequences is an important step in the analysis of duodenogastric reflux (DGR). We propose a joint prediction and segmentation method (JPS) which combines optical flow with active contour to track pylorus. The goal of the proposed method is to improve the pyloric tracking accuracy by taking account of not only the connection information among edge points but also the spatio-temporal information among consecutive frames. The proposed method is compared with other four tracking methods by using both synthetic and real ultrasonic image sequences. Several numerical indexes: Hausdorff distance (HD), average distance (AD), mean edge distance (MED), and edge curvature (EC) have been calculated to evaluate the performance of each method. JPS achieves the minimum distance metrics (HD, AD, and MED) and a smaller EC. The experimental results indicate that JPS gives a better tracking performance than others by the best agreement with the gold curves while keeping the smoothness of the result.

Accurate estimation of fetal weight before delivery is of great benefit to limit the potential complication associated with the low-birth-weight infants. Although the regression analysis has been used as a daily clinical means to estimate the fetal weight on the basis of ultrasound measurements, it still lacks enough accuracy for low-birth-weight fetuses. The ineffectiveness is mainly due to the large inter- or intraobserver variability in measurements and the inappropriateness of the regression analysis. A novel method based on the support vector regression (SVR) is proposed to improve the weight estimation accuracy for fetuses of less than 2500 g. Here, fuzzy logic is introduced into SVR (termed FSVR) to limit the contribution of inaccurate training data to the model establishment, and thus, to enhance the robustness of FSVR to noisy data. To guarantee the generalization performance of the FSVR model, the nondominated sorting genetic algorithm (NSGA) is utilized to obtain the optimal parameters for the FSVR, which is referred to as the evolutionary fuzzy support vector regression (EFSVR) model. Compared with regression formulas, back-propagation neural network, and SVR, EFSVR achieves the lowest mean absolute percent error (6.6%) and the highest correlation coefficient (0.902) between the estimated fetal weight and the actual birth weight. The EFSVR model produces significant improvement (1.9%-4.2%) on the accuracy of fetal weight estimation over several widely used formulas. Experiments show the potential of EFSVR in clinical prenatal care.