Publications
Results
- Showing results for:
- Reset all filters
Search results
-
Journal articleHasan MK, Yang G, Yap CH, 2026,
An efficient, scalable, and adaptable plug-and-play temporal attention module for motion-guided cardiac segmentation with sparse temporal labels
, Medical Image Analysis, Vol: 110, ISSN: 1361-8415Cardiac anatomy segmentation is essential for clinical assessment of cardiac function and disease diagnosis to inform treatment and intervention. Deep learning (DL) has improved cardiac anatomy segmentation accuracy, especially when information on cardiac motion dynamics is integrated into the networks. Several methods for incorporating motion information have been proposed; however, existing methods are not yet optimal: adding the time dimension to input data causes high computational costs, and incorporating registration into the segmentation network remains computationally costly and can be affected by errors of registration, especially with non-DL registration. While attention-based motion modeling is promising, suboptimal design constrains its capacity to learn the complex and coherent temporal interactions inherent in cardiac image sequences. Here, we propose a novel approach to incorporating motion information in the DL segmentation networks: a computationally efficient yet robust Temporal Attention Module (TAM), modeled as a small, multi-headed, cross-temporal attention module, which can be plug-and-play inserted into a broad range of segmentation networks (CNN, transformer, or hybrid) without a drastic architecture modification. Extensive experiments on multiple cardiac imaging datasets, such as 2D echocardiography (CAMUS and EchoNet-Dynamic), 3D echocardiography (MITEA), and 3D cardiac MRI (ACDC), confirm that TAM consistently improves segmentation performance across datasets when added to a range of networks, including UNet, FCN8s, UNetR, SwinUNetR, and the recent I<sup>2</sup>UNet and DT-VNet. Integrating TAM into SAM yields a temporal SAM that reduces Hausdorff distance (HD) from 3.99 mm to 3.51 mm on the CAMUS dataset, while integrating TAM into a pre-trained MedSAM reduces HD from 3.04 to 2.06 pixels after fine-tuning on the EchoNet-Dynamic dataset. On the ACDC 3D dataset, our TAM-UNet and TAM-DT-VNet achieve substantial reductions in HD
-
Journal articleZhou T, Li M, Ruan S, et al., 2026,
A reliable framework for brain tumor segmentation via multi-modal fusion and uncertainty modeling
, Information Fusion, Vol: 129, ISSN: 1566-2535Accurate brain tumor segmentation from MRI scans is critical for effective diagnosis and treatment planning. Recent advances in deep learning have significantly improved brain tumor segmentation performance. However, these models still face challenges in clinical adoption due to their inherent uncertainties and potential for errors. In this paper, we propose a novel MR brain tumor segmentation approach that integrates multi-modal data fusion and uncertainty quantification to improve the accuracy and reliability of brain tumor segmentation. Recognizing that each MR modality contributes unique insights into the tumor’s characteristics, we propose a novel modality-aware guidance by explicitly categorizing the modalities into ”teacher” (FLAIR and T1c) and ”student” (T2 and T1) groups. Since the teacher modalities are the most informative modalities for identifying brain tumors, we propose a multi-modal teacher-student fusion strategy. This strategy leverages the teacher modalities to guide the student modalities in both spatial and channel feature representation aspects. To address prediction reliability, we employ Monte Carlo dropout during training to generate multiple uncertainty estimates. Additionally, we develop a novel uncertainty-aware loss function that optimizes segmentation accuracy while quantifying the uncertainty in predictions. Experimental results conducted on three BraTS datasets demonstrate the effectiveness of the proposed components and the superior performance compared to the state-of-the-art methods, highlighting their potential for clinical application.
-
Journal articleLiao Y, Zheng Y, Zhu J, et al., 2026,
Self-attention-based mixture-of-experts framework for non-invasive prediction of MGMT promoter methylation in glioblastoma using multi-modal MRI
, Displays, Vol: 92, ISSN: 0141-9382Glioblastoma (GBM) is an aggressive brain tumor associated with poor prognosis and limited treatment options. The methylation status of the O6-methylguanine-DNA methyltransferase (MGMT) promoter is a critical biomarker for predicting the efficacy of temozolomide chemotherapy in GBM patients. However, current methods for determining MGMT promoter methylation, including invasive and costly techniques, hinder their widespread clinical application. In this study, we propose a novel non-invasive deep learning framework based on a Mixture-of-Experts (MoE) architecture for predicting MGMT promoter methylation status using multi-modal magnetic resonance imaging (MRI) data. Our MoE model incorporates modality-specific expert networks built on the ResNet18 architecture, with a self-attention-based gating mechanism that dynamically selects and integrates the most relevant features across MRI modalities (T1-weighted, contrast-enhanced T1, T2-weighted, and fluid-attenuated inversion recovery). We evaluate the proposed framework on the BraTS2021 and TCGA-GBM datasets, showing superior performance compared to conventional deep learning models in terms of accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). Furthermore, Grad-CAM visualizations provide enhanced interpretability by highlighting biologically relevant regions in the tumor and peritumoral areas that influence model predictions. The proposed framework represents a promising tool for integrating imaging biomarkers into precision oncology workflows, offering a scalable, cost-effective, and interpretable solution for non-invasive MGMT methylation prediction in GBM.
-
Journal articleSveinsson B, Vangel M, Rowe OE, et al., 2026,
Case Series: Feasibility of Longitudinal Assessment of the Sciatic Nerve in CMT1A Using High-Resolution 7T MRI.
, Muscle NerveINTRODUCTION/AIMS: There is limited data on the sensitivity and responsiveness of high-resolution imaging techniques in the longitudinal assessment of hereditary neuropathies. In this study, our aims were to investigate the ability of ultra-high field magnetic resonance imaging to detect longitudinal changes in the peripheral nerves of Charcot-Marie-Tooth (CMT) 1A patients, and to evaluate the potential benefits of doing so at the nerve fascicle level. METHODS: We performed magnetic resonance imaging (MRI) to simultaneously obtain high-resolution anatomical and quantitative data at ultra-high 7 Tesla field strength in peripheral nerves of four patients with CMT1A disease at baseline and follow up. We compared the resulting measurements of T2 in sciatic, tibial, and fibular nerves within individual fascicles of the three nerve regions. RESULTS: Analyzing individual fascicle distributions, we demonstrated a significantly elevated T2 in the fibular nerve over the course of the study, with a mean increase of 3.55 ms (p = 0.01). Changes in the sciatic nerve were marginally significant (mean increase 1.42 ms, p = 0.05), and tibial nerve changes were not significant (mean increase 1.31 ms, p = 0.18). Combining fascicles across subjects showed significant changes in all three nerves over time. DISCUSSION: Our results indicate that longitudinal MRI assessment of individual nerve fascicles may serve as a quantitative biomarker of disease progression in patients with hereditary neuropathies. Further, our study demonstrates that the data provided by fascicle-level analysis may provide better analytical abilities than whole-nerve imaging.
-
Journal articleZhang C, Wu Y, Boyer-Chammard J, et al., 2026,
Multi-Scale Signal-Image Fusion Model Based On ECoGfor Automatic Detection of Early-stage Traumatic Brain Injury.
, IEEE Trans Biomed Eng, Vol: PPSpreading depolarizations (SDs) are key drivers of secondary brain injury, yet existing bedside monitoring methods that use electrocorticography (ECoG) analyze electrodes and frequency bands separately, thereby obscuring the joint spatiotemporal patterns of SDs. Therefore, this paper introduces a multi-scale signal-image fusion framework that for the first time enables SDmonitoring as a joint multi-modal multi-band spectral image-based analysis. The ECoG signal is converted into a persistent spectral de-weighted spectrogram (PSd-Spec) and joined with multi-band features, through Transformer-CNN jointly empowered blocks: Multi-Channel and Band Transformer Block (MCBTB) and Multi-Scale Adaptive Fusion (MSAF). The network extracts short- and long-range dynamics in a multi-scale time window, while an attention-driven channel weighting module adaptively models the spatial propagation of the electrode strips. On 500h of neuro-ICU recordings, the proposed approach achieved 92.6% accuracy, 84.9% sensitivity. Relative to the best single-modality base line, performance improved by at least 18%, and SD onset was identified on average of 8 min before expert observation. The results suggest that multi-scale fusion of spectral images with ECoG signals yields a clinically actionable early-warning approach and extends quantitative imaging methods to intracranial electrophysiology.
-
Journal articleCheng CW, Huang J, Zhang Y, et al., 2026,
Mamba neural operator: Who wins? transformers vs. state-space models for PDEs
, Journal of Computational Physics, Vol: 548, ISSN: 0021-9991Partial differential equations (PDEs) are widely used to model complex physical systems, but solving them efficiently remains a significant challenge. Recently, Transformers have emerged as the preferred architecture for PDEs due to their ability to capture intricate dependencies. However, they struggle with representing continuous dynamics and long-range interactions. To overcome these limitations, we introduce the Mamba Neural Operator (MNO), a novel framework that enhances neural operator-based techniques for solving PDEs. MNO establishes a formal theoretical connection between structured state-space models (SSMs) and neural operators, offering a unified structure that can adapt to diverse architectures, including Transformer-based models. By leveraging the structured design of SSMs, MNO captures long-range dependencies and continuous dynamics more effectively than traditional Transformers. Through extensive analysis, we show that MNO significantly boosts the expressive power and accuracy of neural operators, making it not just a complement but a superior framework for PDE-related tasks, bridging the gap between efficient representation and accurate solution approximation.
-
Journal articleMa X, Tao Y, Zhang Z, et al., 2026,
Test-time generative augmentation for medical image segmentation
, MEDICAL IMAGE ANALYSIS, Vol: 109, ISSN: 1361-8415 -
Journal articleJing P, Lee K, Zhang Z, et al., 2026,
Reason like a radiologist: Chain-of-thought and reinforcement learning for verifiable report
, MEDICAL IMAGE ANALYSIS, Vol: 109, ISSN: 1361-8415 -
Journal articleAzma Y, Collins D, Lally P, et al., 2026,
Patient-specific biases in fat fraction estimates of malignant bone marrow due to relaxation times measured with STEAM at 3T
, NMR in Biomedicine, ISSN: 0952-3480Semi-quantitative fat fraction estimation using 2-point Dixon sequences is widely used in whole body (WB) MR imaging for malignant bone disease but is biased by relaxation times. Understanding this bias requires water- and fat-specific relaxometry data in normal-appearing marrow and lesions. This study measured bone marrow relaxation times in healthy volunteers and WB-MRI patients using MRS at 3T. Five healthy female volunteers (mean age 38.0 ± 2.5 years) and 24 patients with malignant bone disease undergoing clinical WB-MRI (13 male; mean age 67.7 ± 9.7 years; primary cancers: breast = 5, melanoma = 1, multiple myeloma = 8, prostate = 10) underwent variable inversion/echo time STEAM and 3D gradient echo fat-water imaging. MRS water and fat peaks were fitted to determine T1, T2, R2* (from linewidths), and proton density fat fraction (PDFF). Scan-rescan repeatability of MRS parameters was assessed in volunteers. Lesions were classified by disease state according to clinical reports and segmented in Dixon imaging data for comparison of fat fraction estimates with MRS. Repeatability was evaluated using coefficients of variation. Summary statistics (mean, standard deviation and range) were reported; exploratory inferential statistics were also determined with normality (Shapiro-Wilk) and variance (Levene’s) tests before one-way ANOVA and Tukey’s comparisons (p < 0.05). Monte Carlo simulations assessed relaxation bias on PDFF.All quantitative MRS parameters were repeatable (coefficient of variation < 10%). Water T1 and T2 were most sensitive to disease state in patients, ranging from 1121–2206 ms and 15–71 ms respectively, and were demonstrated to substantially affect 2-point Dixon fat fraction estimates with Monte Carlo simulation. Imaging PDFF achieves closer agreement with MRS PDFF than 2-point Dixon methods. These findings remain preliminary due to the small sample size, but they suggest value in future studies with larger
-
Journal articleLuo Y, Sesia D, Wang F, et al., 2026,
Explicit differentiable slicing and global deformation for cardiac mesh reconstruction.
, Med Image Anal, Vol: 111Three-dimensional (3D) mesh reconstruction of the cardiac anatomy from medical images is useful for shape and motion measurements and biophysics simulations. However, 3D medical images are often acquired as 2D slices that are sparsely sampled (e.g., large slice spacing) and noisy, and 3D mesh reconstruction on such data is a challenging task. Traditional voxel-based approaches utilize non-differentiable pre- and post-processing that compromises fidelity to images, while mesh-level deep learning approaches require large 3D mesh annotations that are difficult to obtain. Differentiable cross-domain supervision from 2D images to 3D meshes is therefore crucial for enabling end-to-end optimization in medical imaging. While there have been attempts to approximate the voxelization and slicing of meshes that are being optimized, there has not yet been a method for directly using 2D slices to supervise 3D mesh reconstruction in a differentiable manner. Here, we propose a novel explicit differentiable voxelization and slicing (DVS) algorithm allowing gradient backpropagation to a 3D mesh from its slices, which facilitates refined mesh optimization directly supervised by the losses defined on 2D images. Further, we propose an innovative framework for extracting patient-specific left ventricle (LV) meshes from medical images by coupling DVS with a graph harmonic deformation (GHD) mesh morphing descriptor of cardiac shape that naturally preserves mesh quality and smoothness during optimization. The proposed framework achieves state-of-the-art performance in cardiac mesh reconstruction tasks from densely sampled (CT) as well as sparsely sampled (MRI stack with few slices) images, outperforming alternatives, including Marching Cubes, statistical shape models, algorithms with vertex-based mesh morphing algorithms and alternative methods for image-supervision of mesh reconstruction. Experimental results demonstrate that our method achieves an overall Dice score of 90% during a sparse
This data is extracted from the Web of Science and reproduced under a licence from Thomson Reuters. You may not copy or re-distribute this data in whole or in part without the written consent of the Science business of Thomson Reuters.