Training object detectors for medical ultrasound videos requires dense frame-by-frame annotations by experts, motivating the use of less costly video-level labels within a weakly semi-supervised object detection (WSSOD) framework. Existing teacher-student approaches typically rely on confidence-based filtering or heuristic spatio-temporal aggregation, which do not provide a principled mechanism for translating global video-level supervision into instance-level pseudo-labels, leading to confirmation bias and error accumulation. We introduce a WSSOD framework that formulates pseudo-label generation as a structured quality assessment problem. Specifically, we explicitly model the relationship between global video labels and local candidate proposals through a graph-based spatiotemporal representation, enabling evaluation of each candidate via long-range relational dependencies. In parallel, candidate proposals are assessed in feature space by measuring their consistency with representations of true-positive and false-positive pathology patterns. By selecting pseudo-labels that are supported by both spatiotemporal context and feature space distributions, the proposed framework effectively translates coarse video-level supervision into reliable instance-level training signals. Experiments on two clinical ultrasound video datasets demonstrate clear quantitative gains over existing SSOD and WSSOD methods using only 10% of frame-level box annotations. On a breast tumor dataset, the proposed method improves over the strongest SSOD baseline by 48% and over the prior WSSOD method by 42%, while recovering 90% of fully supervised performance. On a lung consolidation dataset, the method exceeds the best SSOD and WSSOD baselines by 30% and 17%, respectively, while recovering 75% of fully supervised performance.
Chest X-ray report generation aims to produce diagnostic descriptions grounded in imaging evidence. However, correlation-driven models may exploit reporting templates and disease co-occurrence patterns, leading to unsupported findings and semantic drift. To address this issue, we propose C3E-RRG, a confounder-aware causal debiasing framework for chest X-ray report generation that incorporates explicit causal modeling into the encoder–decoder pipeline. Specifically, we first construct confounding variables by combining language priors derived from pointwise mutual information with disease co-occurrence statistics, and apply confounder-aware adjustment to suppress spurious correlations induced by data bias. We then design an evidence-consistent cross-modal coupling module to strengthen the correspondence between visual features and diagnostic semantics, thereby reducing semantic drift caused by high-frequency reporting templates. Finally, we develop a dynamic bidirectional causal evolution mechanism that iteratively revisits visual evidence and refines semantic representations, progressively correcting semantic deviations and improving image–text consistency during long-sequence generation. Experiments on the IU-XRay and MIMIC-CXR datasets demonstrate competitive performance and validate the effectiveness of the proposed components in suppressing confounding effects, aligning visual evidence with diagnostic semantics, and dynamically correcting generation errors. The source code is available at https://github.com/ cloneiq/C3E-RRG.
Cancer remains a major health concern worldwide. Shear-wave elastography (SWE) holds promise for cancer detection by quantifying linear elastic changes in tissue through shear-wave (SW) velocity estimation. However, shear viscosity and nonlinearity are often over-looked, despite their potential as additional biomarkers. This study evaluates the visco-hyperelastic properties of soft tissue, including linear shear modulus, shear viscosity, and shear nonlinearity. We propose using Volterra kernels to characterize local nonlinear SW dynamics, with kernel identification based on SW particle velocity measurements at adjacent locations. The method was validated using in-silico simulations and in-vitro gelatin phantoms. To evaluate its clinical relevance, we applied the method for prostate cancer (PCa) detection. To our knowledge, this is the first in-vivo assessment of voxel-based visco-hyperelastic properties of the prostate. In 90 patients, our results indicated that the linear kernel correlated with the linear shear modulus, while the quadratic kernel reflected both viscosity and nonlinearity. A convolutional neural network trained on these parameters with registered histopathology as ground truth achieved area under the receiver operating characteristic curve (AUC) values of 0.60 (95% confidence interval [CI]: 0.57, 0.62) and 0.66 (95% CI: 0.64, 0.68) for PCa detection, respectively—both significantly outperforming the linear shear modulus (AUC of 0.54; 95% CI, 0.52 to 0.56). Our findings suggest that Volterra kernels reflect information related to the visco-hyperelastic properties of prostate tissue, highlighting the potential of viscosity and nonlinearity as additional biomarkers for PCa detection.
State space model (SSM) is a powerful tool in neuroscience field to characterize the dynamic nature of brain functions by elucidating the mechanism of how brain system transits between brain states and how underlying states give rise to the observed neural activities. Although tremendous efforts have been made to lend the power of deep learning and mathematical insight of SSM in various functional neuroimaging studies, current state-of-the-art methods lack a holistic view of brain state evolution as a self-organized dynamical system where each part of the brain is functionally inter-connected. Since the topological co-activation of functional fluctuations exhibits an intrinsic geometric pattern (symmetric and positive definite, SPD) on the Riemannian manifold, the call for understanding how a selective set of functional connectivities in the brain supports diverse behavior and cognition emerge a new machine learning scenario of manifold-based SSM for largescale functional neuroimages. To that end, we propose a geometric neural network of SSM, coined GeoMind, designed to uncover evolving brain states by tracking the trajectory of functional dynamics on a high-dimensional Riemannian manifold of SPD matrices. GeoMind demonstrates promising results in identifying specific brain states on task-based functional Magnetic Resonance Imaging (fMRI) data, as well as in diseases early diagnosis for Alzheimer’s disease, Parkinson’s disease and Autism, by using resting-state fMRI. These results highlight the applicability of GeoMind in neuroscience research. To further evaluate generalizability of GeoMind, we applied GeoMind to electroencephalography (EEG) emotion datasets, outperforming competing methods and demonstrating scalability across neuroimaging modalities.
