Results
- Showing results for:
- Reset all filters
Search results
-
Journal articleWei X, Liao Y, Zhu J, et al., 2026,
A hybrid CNN–Mamba state space model with pyramid-pooled skip connections for prostate tumor segmentation
, Expert Systems with Applications, Vol: 327, ISSN: 0957-4174Automatic prostate tumor segmentation in multiparametric magnetic resonance imaging (mpMRI) can support targeted biopsy and therapy planning, but remains challenging due to the small, heterogeneous and low-contrast nature of the lesions. CNN-based encoder-decoder models capture local detail well, but struggle with long-range reasoning. Transformers excel at global interactions, but are computationally expensive. Mamba/state-space models enable efficient global modelling, but often lack boundary-sensitive local refinement. We propose ProMamba: a hybrid CNN-Mamba segmentation framework combining local detail extraction with efficient global dependency modelling within a hierarchical U-Net architecture. Its dual-branch Pro-SSM block incorporates a depthwise dilated convolution module (DDCM) to extract multi-scale boundary cues, as well as a Mamba-based visual state-space (VSS) branch to aggregate multi-directional long-range context. The two branches are fused via channel concatenation and shuffling while preserving the 2D convolutional pathway and avoiding spatial flattening. Pyramid-pooled skip connections further reduce the semantic gap between the encoder and decoder and improve small-lesion delineation. Using five-fold cross-validation at the patient level on Prostate158, PI-CAI 2022 and an ethics-approved private cohort of 110 cases, ProMamba achieves Dice scores of 51.39%, 59.46% and 48.49% respectively, outperforming strong CNN, nnU-Net, efficient Transformer/hybrid and Mamba baselines. ProMamba also shows competitive latency and memory usage and generalizes well to PROMISE12 and ACDC, achieving Dice scores of 91.40% and 92.14%, respectively.
-
Journal articleWen K, Ferreira PF, Di Biase Oemick A, et al., 2026,
Evaluation of Third-Order Motion-Compensated Cardiac Diffusion Tensor Imaging Across Cardiac Phases Using an Ultra-High-Performance Clinical Scanner.
, Magn Reson Med, Vol: 96, Pages: 1365-1378PURPOSE: To evaluate a third-order motion-compensated spin echo (M3-MCSE) sequence at multiple cardiac phases on a clinical 3 T MRI scanner with ultra-high performance (UHP) gradients (200 mT/m), compared with stimulated echo acquisition mode (STEAM) and second-order MCSE (M2-MCSE) for cardiac diffusion tensor imaging (cDTI). METHODS: Twenty healthy subjects underwent mid-ventricular short-axis cDTI at peak systole and diastasis using STEAM, M2-MCSE, and M3-MCSE. cDTI metrics and image quality were compared. In five additional healthy subjects, diffusion-weighted images were obtained at multiple trigger delays distributed over diastasis to assess motion-induced signal loss. RESULTS: Compared to M2-MCSE, M3-MCSE yielded higher systolic helix angle map scores ( p = 0.007 ) but lower diastolic scores ( p = 0.001 ), with no significant difference in mean diffusivity, fractional anisotropy, helix angle transmurality or sheetlet angle in systole/diastole. STEAM-derived apparent diffusion coefficients (ADC) were consistent across diastasis, while ADC for MCSE sequences increased at sub-optimal trigger delays. CONCLUSION: UHP gradients enabled in vivo evaluation of M3-MCSE, showing superior systolic cDTI but reduced diastolic performance versus M2-MCSE due to reduced signal-to-noise ratio and a longer motion-sensitive window. Future work may consider numerically optimized gradient designs to enhance MCSE robustness throughout the cardiac cycle.
-
Journal articleLuo Y, Ferreira PF, Wen K, et al., 2026,
Optimized Reduced Field of View and Fat Suppression Methods for Interleaved Multislice In Vivo Cardiac Diffusion Tensor Imaging.
, Magn Reson Med, Vol: 96, Pages: 1097-1110PURPOSE: Slice interleaving, a limited phase encode (PE) field of view (FOV), and effective fat suppression are vital for efficient cardiac diffusion tensor imaging (cDTI) with minimal artifacts. This study aimed to optimize reduced FOV and fat suppression methods for interleaved multislice cDTI to improve signal-to-noise ratio (SNR) and minimize artifacts. METHODS: Two-slice motion compensated spin echo datasets from 20 healthy volunteers were acquired. Four reduced PE FOV sequences were evaluated: 2DRF pulse; applying either 180 ° or 90 ° pulses in PE direction; and the proposed flip-back sequence with a nonselective 180 ° pulse after readout to restore inverted magnetization. Four fat suppression techniques were implemented: no fat suppression (standard); fat saturation; binomial water excitation and spectral attenuated inversion recovery (SPAIR). RESULTS: The proposed flip-back sequence with SPAIR achieved the highest median SNR, and its SNR values are significantly higher ( p < 0.01 ) than 2DRF with SPAIR as current state-of-the-art. SPAIR and water excitation demonstrated comparable performance when combined with the flip-back sequence, and both yielded superior image quality than with no suppression or fat saturation. SPAIR showed robust fat suppression across most subjects, whilst water excitation exhibited advantages in some subjects with a high body mass index. CONCLUSION: The proposed flip-back sequence with SPAIR enables efficient interleaved multislice imaging with reduced PE FOV and effective fat suppression, facilitating clinical translation of in vivo cDTI.
-
Journal articleDeng F, Chu X, Shi W, et al., 2026,
Radiogenomic modeling of EGFR mutation status in brain metastases from lung adenocarcinoma: a multicenter study with biological interpretability
, npj Digital Medicine, ISSN: 2398-6352 -
Journal articleNonchev K, Andani S, Ficek-Pascual J, et al., 2026,
Representation learning for multi-modal spatially resolved transcriptomics data.
, Bioinformatics, Vol: 42MOTIVATION: Spatial transcriptomics enables in-depth molecular characterization of samples on a morphology and RNA level while preserving spatial location. Integrating the resulting multi-modal data is an unsolved problem, and developing new solutions in precision medicine depends on improved methodologies. RESULTS: We introduce AESTETIK, a convolutional deep learning model that jointly integrates spatial, transcriptomics, and morphology information to learn accurate spot representations. AESTETIK yielded substantially improved cluster assignments on widely adopted technology platforms (e.g. 10x Genomics™, NanoString™) across multiple datasets. We achieved performance enhancement on structured tissues (e.g. brain) with a 21% increase in median ARI over previous state-of-the-art methods. Notably, AESTETIK also demonstrated superior performance on cancer tissues with heterogeneous cell populations, showing a 2-fold increase in breast cancer, 79% in melanoma, and 21% in liver cancer. We expect that these advances will enable a multi-modal understanding of key biological processes. AVAILABILITY AND IMPLEMENTATION: AESTETIK is implemented in Python 3 and is available as open source software at http://www.github.com/ratschlab/aestetik. The Snakemake pipeline for reproducing the results is available at http://www.github.com/ratschlab/st-rep.
-
Journal articleLee K, Jing P, Zhang Z, et al., 2026,
Seeing through experts' eyes: a foundational vision-language model trained on radiologists' gaze and reasoning
, npj Artificial Intelligence, ISSN: 3005-1460Large-scale vision–language models have shown promise in automating chest X-ray interpretation. However, their clinical utility remains limited by a fundamental gap between model outputs and the reasoning processes of radiologists. Most systems optimize for semantic information without emulating how experts visually examine and interpret medical images. As a result, current models often overlook critical findings, misrepresent anatomical context, or generate reports that diverge from established diagnostic workflows. Radiologists typically follow structured protocols (e.g., the ABCDEF approach that sequentially assesses the airways, breathing, circulation, diaphragm, external and foreign material). This standardized workflow ensures that all clinically relevant regions are examined in a consistent and systematic manner. It reduces the risk of missed findings, supports reliable diagnostic reasoning and facilitates clear communication between radiologists and referring clinicians. Emulating such expert strategies is essential for improving the trustworthiness and interpretability of automated report generation systems. Therefore, we introduce Gaze-X, a vision–language model that leverages radiologists’ eye-tracking data as a behavioral prior to model expert diagnostic reasoning. By incorporating gaze trajectories and fixation patterns into pretraining, Gaze-X learns to follow the spatial and temporal structure of radiologist attention and integratesvisual observations in a clinically meaningful sequence. This approach enables the model to align itsfocus with diagnostically relevant regions and to emulate the interpretive logic underlying expert reports. Using a curated dataset of gaze recordings, comprising over 30,000 key frames across diversedisease categories, from five radiologists interpreting chest X-rays, we demonstrate that Gaze-X produces more accurate, interpretable, and expert-consistent outputs across a range of clinically relevanttasks a
-
Journal articleSedlacik J, McGurk KA, Tokarczuk PF, et al., 2026,
Quantitative ventricular trabeculation assessment in cardiac MRI: optimised blood-pool segmentation, box-counting fractal analysis and non-fractal measurements.
, Int J Cardiovasc Imaging, Vol: 42, Pages: 1369-1379Quantitative assessment of the ventricular trabeculation by fractal dimension (FD) involves complex processing steps which may impact the results. We optimised the automated processing workflow for a reliable assessment of the left and right ventricles at end-diastole and end-systole which is suitable for the automated analysis of large-scale cohorts. Ventricular trabeculae and blood were segmented using a level-set method optimised to exclude pixels outside the heart on short-axis cardiac MRI. FD was derived by box-counting the trabeculae/blood boundary while investigating the impact of box size, sampling and rotation. Alternative non-fractal measures - the convexity related boundary length ratio (BLR) and the trabeculated mass ratio (TMR) - were also investigated.FD values with and without optimisation showed a strong linear correlation (R2 = 0.81) and narrow agreement limit (1.96·SD = 0.063) only for the end-diastolic left ventricle. Linear correlation and agreement was good between the optimised FD and BLR values for both ventricles and cardiac phases (R2 = 0.70-0.92, 1.96·SD = 0.037-0.064) but not for TMR (R2 = 0-0.37, 1.96·SD = 0.16-1.4). FD, BLR and TMR differed significantly (p < 0.001) between end-diastole and end-systole with lower FD (-0.07 ± 0.06) but higher BLR (0.31 ± 0.25) and TMR (0.26 ± 0.13) values at end-systole.The previously used fractal analysis is suboptimal except for assessing the end-diastole left ventricle. The optimised fractal analysis is suitable for the left and right ventricle at end-diastole and end-systole. The easy to compute non-fractal BLR gives equivalent information like FD. The volume-based TMR, on the other hand, captures different features of the trabeculation.
-
Journal articleTänzer M, Lim EJ, Qiu HH, et al., 2026,
Simultaneous multi-slice Cardiac Diffusion Tensor Imaging with variable CAIPIRINHA shifts and artefact-aware AI.
, Med Image Anal, Vol: 112Cardiac Diffusion Tensor Imaging (cDTI) provides unique insights into myocardial microstructure in-vivo but requires averaging multiple repetitions for adequate signal quality, leading to prohibitively long acquisition times. Standard acceleration strategies, such as reducing repetitions and employing simultaneous multi-slice (SMS) imaging, are limited by low signal-to-noise ratio (SNR) and inter-slice leakage artefacts, respectively. We introduce ORCAS, a unified framework that synergistically combines a novel variable CAIPIRINHA acquisition with an artefact-aware AI reconstruction to overcome these challenges. The variable CAIPIRINHA scheme decoheres SMS artefacts across repetitions, while our dual-domain deep learning model simultaneously suppresses these artefacts and combats the low SNR from fewer repetitions. The model is guided by patient-specific single-band auxiliary data to preserve anatomical fidelity. Validated on ex-vivo hearts with and without anomalies, ORCAS achieves an over 18-fold acceleration by combining these strategies, reducing a whole-heart scan from over two hours to under 7 min. This is accomplished while reducing errors in key biomarkers, such as Fractional Anisotropy, by up to 64%. The framework preserves essential microstructural properties and the delineation of abnormalities, representing a significant step towards the clinical translation of whole-heart cDTI.
-
Journal articleWang T, Zhang Z, Zhou Y, et al., 2026,
From noisy labels to intrinsic structure: A geometric-structural dual-guided framework for noise-robust medical image segmentation.
, Med Image Anal, Vol: 112The effectiveness of convolutional neural networks in medical image segmentation relies on large-scale, high-quality annotations, which are costly and time-consuming to obtain. Even expert-labeled datasets inevitably contain noise arising from subjectivity and coarse delineations, which disrupt feature learning and adversely impact model performance. To address these challenges, this study proposes a Geometric-Structural Dual-Guided Network (GSD-Net), which integrates geometric and structural cues to improve robustness against noisy annotations. It incorporates a Geometric Distance-Aware module that dynamically adjusts pixel-level weights using geometric features, thereby strengthening supervision in reliable regions while suppressing noise. A Structure-Guided Label Refinement module further refines labels with structural priors, and a Knowledge Transfer module enriches supervision and improves sensitivity to local details. To comprehensively assess its effectiveness, we evaluated GSD-Net on six publicly available datasets: four containing three types of simulated label noise, and two with multi-expert annotations that reflect real-world subjectivity and labeling inconsistencies. Experimental results demonstrate that GSD-Net achieves state-of-the-art performance under noisy annotations, achieving improvements of 1.58% on Kvasir, 22.76% on Shenzhen, 8.87% on BU_SUC, and 1.77% on BraTS2020 under SR simulated noise. The code of this study is available at https://github.com/ortonwang/GSD-Net.
-
Journal articleZhang Z, Jing P, Wang Z, et al., 2026,
Cyclic Self-Supervised Diffusion for Ultra Low-Field to High-Field MRI Synthesis.
, IEEE Trans Med Imaging, Vol: 45, Pages: 3766-3777Synthesizing high-quality images from low-field MRI holds significant potential. Low-field MRI is cheaper, more accessible, and safer, but suffers from low resolution and poor signal-to-noise ratio. This synthesis process can reduce reliance on costly acquisitions and expand data availability. However, synthesizing high-field MRI still suffers from a clinical fidelity gap. There is a need to preserve anatomical fidelity, enhance fine-grained structural details, and bridge domain gaps in image contrast. To address these issues, we propose a cyclic self-supervised diffusion (CSS-Diff) framework for high-field MRI synthesis from real low-field MRI data. Our core idea is to reformulate diffusion-based synthesis under a cycle-consistent constraint. It enforces anatomical preservation throughout the generative process rather than just relying on paired pixel-level supervision. The CSS-Diff framework further incorporates two novel processes. The slice-wise gap perception network aligns inter-slice inconsistencies via contrastive learning. The local structure correction network enhances local feature restoration through self-reconstruction of masked and perturbed patches. Extensive experiments on cross-field synthesis tasks demonstrate the effectiveness of our method, achieving state-of-the-art performance (e.g., $31.80~\pm ~2.70$ dB in PSNR, $0.943~\pm ~0.102$ in SSIM, and $0.0864~\pm ~0.0689$ in LPIPS). Beyond pixel-wise fidelity, our method also preserves fine-grained anatomical structures compared with the original low-field MRI (e.g., left cerebral white matter error drops from 12.1% to 2.1%, cortex from 4.2% to 3.7%). To conclude, our CSS-Diff can synthesize images that are both quantitatively reliable and anatomically consistent. The code is available at: https://github.com/ayanglab/CSS-Diff.
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.
Contact
For enquiries about the MRI Physics Collective, please contact:
Mary Finnegan
Senior MR Physicist at the Imperial College Healthcare NHS Trust
Pete Lally
Assistant Professor in Magnetic Resonance (MR) Physics at Imperial College
Jan Sedlacik
MR Physicist at the Robert Steiner MR Unit, Hammersmith Hospital Campus