ACS Sens. 2025 Mar 23. doi: 10.1021/acssensors.5c00254. Online ahead of print.

ABSTRACT

Various sensing methods have been developed for food spoilage research, but in practical applications, the accuracy of these methods is frequently constrained by the limitation of single-source data and challenges in cross-validating multimodal data. To address these issues, a new method combining multidimensional sensing technology with deep learning-based dynamic fusion has been developed, which can precisely monitor the spoilage process of beef. This study designs a gas sensor based on surface-enhanced Raman scattering (SERS) to directly analyze volatile organic compounds (VOCs) adsorbed on MIL-101(Cr) with amine-specific adsorption for data collection while also evaluating the moisture distribution of beef through low-field nuclear magnetic resonance (LF-NMR), providing multidimensional recognition and readings. By introducing the self-attention mechanism and SENet scaling features into the multimodal deep learning model, the system is able to adaptively fuse and focus on the important features of the sensors. After training, the system can predict the storage time of beef under controlled storage conditions, with an R2 value greater than 0.98. Furthermore, it can provide accurate freshness assessments for beef samples under unknown storage conditions. Relative to single-modality methods, accuracy improves from 90 to over 97%. Overall, the newly developed dynamic fusion deep learning multimodal model effectively integrates multimodal information, enabling the fast and reliable monitoring of beef freshness.

PMID:40123082 | DOI:10.1021/acssensors.5c00254

Z Med Phys. 2025 Mar 22:S0939-3889(25)00035-2. doi: 10.1016/j.zemedi.2025.02.005. Online ahead of print.

ABSTRACT

OBJECTIVES: High prevalence of visceral obesity and its associated complications underscore the importance of accurately quantifying visceral adipose tissue (VAT) depots. While whole-body MRI offers comprehensive insights into adipose tissue distribution, it is resource-intensive. Alternatively, evaluation of defined single slices provides an efficient approach for estimation of total VAT volume. This study investigates the influence of sex-, age-, and BMI on VAT distribution along the craniocaudal axis and total VAT volume obtained from single slice versus volumetric assessment in 3D MRI and aims to identify age-independent locations for accurate estimation of VAT volume from single slice assessment.

MATERIALS AND METHODS: This secondary analysis of the prospective population-based German National Cohort (NAKO) included 3D VIBE Dixon MRI from 11,191 participants (screened between May 2014 and December 2016). VAT and spine segmentations were automatically generated using fat-selective images. Standardized craniocaudal VAT profiles were generated. Axial percentage of total VAT was used for identification of reference locations for volume estimation of VAT from a single slice.

RESULTS: Data from 11,036 participants (mean age, 52 ± 11 years, 5681 men) were analyzed. Craniocaudal VAT distribution differed qualitatively between men/women and with respect to age/BMI. Age-independent single slice VAT estimates demonstrated strong correlations with reference VAT volumes. Anatomical locations for accurate VAT estimation varied with sex/BMI.

CONCLUSIONS: The selection of reference locations should be different depending on BMI groups, with a preference for caudal shifts in location with increasing BMI. For women with obesity (BMI >30 kg/m2), the L1 level emerges as the optimal reference location.

PMID:40122750 | DOI:10.1016/j.zemedi.2025.02.005

Heart. 2025 Mar 23:heartjnl-2024-323947. doi: 10.1136/heartjnl-2024-323947. Online ahead of print.

ABSTRACT

Cardiac telemetry has evolved into a vital tool for continuous cardiac monitoring and early detection of cardiac abnormalities. In recent years, artificial intelligence (AI) has become increasingly integrated into cardiac telemetry, making a shift from traditional statistical machine learning models to more advanced deep neural networks. These modern AI models have demonstrated superior accuracy and the ability to detect complex patterns in telemetry data, enhancing real-time monitoring, predictive analytics and personalised cardiac care. In our review, we examine the current state of AI in cardiac telemetry, focusing on deep learning techniques, their clinical applications, the challenges and limitations faced by these models, and potential future directions in this promising field.

PMID:40122590 | DOI:10.1136/heartjnl-2024-323947

Neuroimage. 2025 Mar 21:121145. doi: 10.1016/j.neuroimage.2025.121145. Online ahead of print.

ABSTRACT

Magnetic resonance imaging (MRI)-derived hippocampus measurements have been associated with different cognitive domains. The knowledge of hippocampal structural deformations as we age has contributed to our understanding of the overall aging process. Different morphological hippocampal shape analysis methods have been developed, but it is unclear how their principles relate and how consistent are the published results in relation to cognition in the normal elderly in the light of the new deep-learning-based (DL) state-of-the-art modeling methods. We compared results from analyzing the hippocampal morphology using manually-generated binary masks and a Laplacian- based deformation shape analysis method, with those resulting from analyzing SynthSeg-generated hippocampal binary masks using a DL method based on the PointNet architecture, in relation to different cognitive domains. Whilst most previously reported statistically significant associations were also replicated, differences were also observed due to 1) differences in the binary masks and 2) differences in sensitivity between the methods. Differences in the template mesh, number of vertices of the template mesh, and their distribution did not impact the results.

PMID:40122476 | DOI:10.1016/j.neuroimage.2025.121145

J Neurosci Methods. 2025 Mar 21:110429. doi: 10.1016/j.jneumeth.2025.110429. Online ahead of print.

ABSTRACT

BACKGROUND: The Lateral Geniculate Nucleus (LGN) is an essential contributor to high-level visual processing despite being an early subcortical area in the visual system. Current LGN computational models focus on its basic properties, with less emphasis on its role in high-level vision.

NEW METHOD: We propose a high-level approach for encoding mouse LGN neural responses to natural scenes. This approach employs two deep neural networks (DNNs); namely VGG16 and ResNet50, as goal-driven models. We use these models as tools to better understand visual features encoded in the LGN.

RESULTS: Early layers of the DNNs represent the best LGN models. We also demonstrate that numerosity, as a high-level visual feature, is encoded, along with other visual features, in LGN neural activity. Results demonstrate that intermediate layers are better in representing numerosity compared to early layers. Early layers are better at predicting simple visual features, while intermediate layers are better at predicting more complex features. Finally, we show that an ensemble model of an early and an intermediate layer achieves high neural prediction accuracy and numerosity representation.

COMPARISON WITH EXISTING METHOD(S): Our approach emphasizes the role of analyzing the inner workings of DNNs to demonstrate the representation of a high-level feature such as numerosity in the LGN, as opposed to the common belief about the simplicity of the LGN.

CONCLUSIONS: We demonstrate that goal-driven DNNs can be used as high-level vision models of the LGN for neural prediction and as an exploration tool to better understand the role of the LGN.

PMID:40122470 | DOI:10.1016/j.jneumeth.2025.110429

J Neurosci Methods. 2025 Mar 21:110424. doi: 10.1016/j.jneumeth.2025.110424. Online ahead of print.

ABSTRACT

BACKGROUND: Accurate brain tumor segmentation from MRI images is critical in the medical industry, directly impacts the efficacy of diagnostic and treatment plans. Accurate segmentation of tumor region can be challenging, especially when noise and abnormalities are present.

METHOD: This research provides a systematic review of automatic brain tumor segmentation techniques, with a specific focus on the design of network architectures. The review categorizes existing methods into unsupervised and supervised learning techniques, as well as machine learning and deep learning approaches within supervised techniques. Deep learning techniques are thoroughly reviewed, with a particular focus on CNN-based, U-Net-based, transfer learning-based, transformer-based, and hybrid transformer-based methods.

SCOPE AND COVERAGE: This survey encompasses a broad spectrum of automatic segmentation methodologies, from traditional machine learning approaches to advanced deep learning frameworks. It provides an in-depth comparison of performance metrics, model efficiency, and robustness across multiple datasets, particularly the BraTS dataset. The study further examines multi-modal MRI imaging and its influence on segmentation accuracy, addressing domain adaptation, class imbalance, and generalization challenges.

COMPARISON WITH EXISTING METHODS: The analysis highlights the current challenges in Computer-aided Diagnostic (CAD) systems, examining how different models and imaging sequences impact performance. Recent advancements in deep learning, especially the widespread use of U-Net architectures, have significantly enhanced medical image segmentation. This review critically evaluates these developments, focusing the iterative improvements in U-Net models that have driven progress in brain tumor segmentation. Furthermore, it explores various techniques for improving U-Net performance for medical applications, focussing on its potential for improving diagnostic and treatment planning procedures.

CONCLUSION: The efficiency of these automated segmentation approaches is rigorously evaluated using the BraTS dataset, a benchmark dataset, part of the annual Multimodal Brain Tumour Segmentation Challenge (MICCAI). This evaluation provides insights into the current state-of-the-art and identifies key areas for future research and development.

PMID:40122469 | DOI:10.1016/j.jneumeth.2025.110424

Hum Pathol. 2025 Mar 21:105762. doi: 10.1016/j.humpath.2025.105762. Online ahead of print.

ABSTRACT

Computational pathology has primarily focused on analyzing tissue slides, neglecting the valuable information contained in gross images. To bridge this gap, we proposed a novel approach leveraging the Swin Transformer architecture to develop a Swin-Transformer based Gross Features Detective Network (SGFD-network), which assist pathologists for locating diseased area in ovarian epithelial tumors based on their gross features. Our model was trained on 4129 gross images and achieved high accuracy rates of 88.9%, 86.4%, and 93.0% for benign, borderline, and carcinoma group, respectively, demonstrating strong agreement with pathologist evaluations. Notably, we trained a new classifier to distinguish between borderline tumors and those with microinvasion or microinvasive carcinoma, addressing a significant challenge in frozen section sampling. Our study was the first to propose a solution to this challenge, showcasing high accuracy rates of 85.0% and 92.2% for each group, respectively. To further elucidate the decision-making process, we employed Class Activation Mapping-grad to identify high-contribution zones and applied k-means clustering to summarize these features. The resulting clustered features can effectively complement existing knowledge of gross examination, improving the distinction between borderline tumors and those with microinvasion or microinvasive carcinoma. Our model identifies high-risk areas for microinvasion or microinvasive carcinoma, enabling pathologists to target sampling more effectively during frozen sections. Furthermore, SGFD-network requires only a single 4090 graphics card and completes a single interpretation task in 3 minutes. This study demonstrates the potential of deep learning-based analysis of gross features to aid in ovarian epithelial tumors sampling, especially in frozen section.

PMID:40122402 | DOI:10.1016/j.humpath.2025.105762

J Cardiovasc Magn Reson. 2025 Mar 21:101886. doi: 10.1016/j.jocmr.2025.101886. Online ahead of print.

ABSTRACT

BACKGROUND: Patients with repaired tetralogy of Fallot (rTOF) are commonly followed with MRI and frequently develop right ventricular (RV) dysfunction, which can be severe enough to impact left ventricular (LV) function in some patients. In this study, we sought to characterize patterns of LV dysfunction in this patient population using Deep Learning Synthetic Strain (DLSS), a fully automated deep learning algorithm capable of measuring regional LV strain and dyssynchrony.

METHODS: We retrospectively collected cine SSFP MRI images from a multi-institutional cohort of 198 patients with rTOF and 21 healthy controls. Using DLSS, we measured LV strain and strain rate across 16 AHA segments from short-axis cine SSFP images and compared these values to controls. We then performed a clustering analysis to identify unique patterns of LV contraction, using segmental peak strain and several measures of dyssynchrony. We further characterized these patterns by assessing their relationship to traditional MRI metrics of volume and function. Lastly, we assessed their impact on subsequent progression to pulmonary valve replacement (PVR) through a multivariate analysis.

RESULTS: Overall, patients with rTOF had decreased septal radial strain, increased lateral wall radial strain, and increased dyssynchrony relative to healthy controls. Clustering of rTOF patients identified four unique patterns of LV contraction. Most notably, patients in cluster 1 (n=39) demonstrated an LV contraction pattern with paradoxical septal wall motion and severely reduced septal strain. These patients had significantly elevated RV end-diastolic volume relative to clusters 3 and 4 (153±34 vs. 127±34 and 126±31mL/m2, ANOVA p<0.01). In the multivariate analysis, this contraction pattern was the only LV metric associated with future progression to pulmonary valve replacement (HR = 2.69, p<0.005). A smaller subset of patients (cluster 2, n=29) showed reduced septal strain and LV ejection fraction despite synchronous ventricular contraction.

CONCLUSIONS: Patients with rTOF demonstrate four unique patterns of LV dysfunction. Most commonly, but not exclusively, LV dysfunction is characterized by septal wall motion abnormalities and severely reduced septal strain. Patients with this pattern of LV dysfunction had concomitant RV dysfunction and rapid progression to PVR.

PMID:40122390 | DOI:10.1016/j.jocmr.2025.101886

Magn Reson Imaging. 2025 Mar 21:110386. doi: 10.1016/j.mri.2025.110386. Online ahead of print.

ABSTRACT

Medical image registration plays a crucial role in medical imaging, with a wide range of clinical applications. In this context, brain MRI registration is commonly used in clinical practice for accurate diagnosis and treatment planning. In recent years, deep learning-based deformable registration methods have achieved remarkable results. However, existing methods have not been flexible and efficient in handling the feature relationships of anatomical structures at different levels when dealing with large deformations. To address this limitation, we propose a novel strategic integration registration network based on the pyramid structure. Our strategy mainly includes two aspects of integration: fusion of features at different scales, and integration of different neural network structures. Specifically, we design a CNN encoder and a Transformer decoder to efficiently extract and enhance both global and local features. Moreover, to overcome the error accumulation issue inherent in pyramid structures, we introduce progressive optimization iterations at the lowest scale for deformation field generation. This approach more efficiently handles the spatial relationships of images while improving accuracy. We conduct extensive evaluations across multiple brain MRI datasets, and experimental results show that our method outperforms other deep learning-based methods in terms of registration accuracy and robustness.

PMID:40122188 | DOI:10.1016/j.mri.2025.110386

EBioMedicine. 2025 Mar 22;114:105663. doi: 10.1016/j.ebiom.2025.105663. Online ahead of print.

ABSTRACT

BACKGROUND: We aim to predict outcomes of human papillomavirus (HPV)-associated oropharyngeal squamous cell carcinoma (OPSCC), a subtype of head and neck cancer characterized with improved clinical outcome and better response to therapy. Pathology and radiology focused AI-based prognostic models have been independently developed for OPSCC, but their integration incorporating both primary tumour (PT) and metastatic cervical lymph node (LN) remains unexamined.

METHODS: We investigate the prognostic value of an AI approach termed the swintransformer-based multimodal and multi-region data fusion framework (SMuRF). SMuRF integrates features from CT corresponding to the PT and LN, as well as whole slide pathology images from the PT as a predictor of survival and tumour grade in HPV-associated OPSCC. SMuRF employs cross-modality and cross-region window based multi-head self-attention mechanisms to capture interactions between features across tumour habitats and image scales.

FINDINGS: Developed and tested on a cohort of 277 patients with OPSCC with matched radiology and pathology images, SMuRF demonstrated strong performance (C-index = 0.81 for DFS prediction and AUC = 0.75 for tumour grade classification) and emerged as an independent prognostic biomarker for DFS (hazard ratio [HR] = 17, 95% confidence interval [CI], 4.9-58, p < 0.0001) and tumour grade (odds ratio [OR] = 3.7, 95% CI, 1.4-10.5, p = 0.01) controlling for other clinical variables (i.e., T-, N-stage, age, smoking, sex and treatment modalities). Importantly, SMuRF outperformed unimodal models derived from radiology or pathology alone.

INTERPRETATION: Our findings underscore the potential of multimodal deep learning in accurately stratifying OPSCC risk, informing tailored treatment strategies and potentially refining existing treatment algorithms.

FUNDING: The National Institutes of Health, the U.S. Department of Veterans Affairs and National Institute of Biomedical Imaging and Bioengineering.

PMID:40121941 | DOI:10.1016/j.ebiom.2025.105663

Comput Med Imaging Graph. 2025 Mar 21;123:102532. doi: 10.1016/j.compmedimag.2025.102532. Online ahead of print.

ABSTRACT

Multimodality is often necessary for improving object segmentation tasks, especially in the case of multilabel tasks, such as tumor segmentation, which is crucial for clinical diagnosis and treatment planning. However, a major challenge in utilizing multimodality with deep learning remains: the limited availability of annotated training data, primarily due to the time-consuming acquisition process and the necessity for expert annotations. Although deep learning has significantly advanced many tasks in medical imaging, conventional augmentation techniques are often insufficient due to the inherent complexity of volumetric medical data. To address this problem, we propose an innovative slice-based latent diffusion architecture for the generation of 3D multi-modal images and their corresponding multi-label masks. Our approach enables the simultaneous generation of the image and mask in a slice-by-slice fashion, leveraging a positional encoding and a Latent Aggregation module to maintain spatial coherence and capture slice sequentiality. This method effectively reduces the computational complexity and memory demands typically associated with diffusion models. Additionally, we condition our architecture on tumor characteristics to generate a diverse array of tumor variations and enhance texture using a refining module that acts like a super-resolution mechanism, mitigating the inherent blurriness caused by data scarcity in the autoencoder. We evaluate the effectiveness of our synthesized volumes using the BRATS2021 dataset to segment the tumor with three tissue labels and compare them with other state-of-the-art diffusion models through a downstream segmentation task, demonstrating the superior performance and efficiency of our method. While our primary application is tumor segmentation, this method can be readily adapted to other modalities. Code is available here : https://github.com/Arksyd96/multi-modal-mri-and-mask-synthesis-with-conditional-slice-based-ldm.

PMID:40121926 | DOI:10.1016/j.compmedimag.2025.102532

Insights Imaging. 2025 Mar 23;16(1):68. doi: 10.1186/s13244-025-01937-3.

ABSTRACT

OBJECTIVES: A novel risk stratification model based on Lung-RADS® v2022 and CT features was constructed and validated for predicting invasive pure ground-glass nodules (pGGNs) in China.

METHODS: Five hundred and twenty-six patients with 572 pulmonary GGNs were prospectively enrolled and divided into training (n = 169) and validation (n = 403) sets. Utilising the Lung-RADS® v2022 framework and the types of GGN-vessel relationships (GVR), a complementary Lung-RADS® v2022 was established, and the pGGNs were reclassified from categories 2, 3 and 4x of Lung-RADS® v2022 into 2, 3, 4a, 4b, and 4x of cLung-RADS® v2022. The cutoff value of invasive pGGNs was defined as the cLung-RADS® v2022 4a-4x. Evaluation metrics like recall rate, precision, F1 score, accuracy, Matthews correlation coefficient (MCC), and the area under the receiver operating characteristic curve (AUC) were employed to assess the utility of the cLung-RADS® v2022.

RESULTS: In the training set, compared with the Lung-RADS 1.0, the AUC of Lung-RADS® v2022 were decreased from 0.543 to 0.511 (p-value = 0.002), and compared to Lung-RADS 1.0 and Lung-RADS® v2022, the cLung-RADS® v2022 model exhibited the highest recall rate (94.9% vs 6.5%, 2.2%), MCC value (60.2% vs 5.4%, 6.3%), F1 score (92.5% vs 12.1%, 4.3%), accuracy (87.6% vs 23.1%, 19.5%), and AUC (0.718 vs 0.543, 0.511; p-value = 0.014, 0.0016) in diagnosing the invasiveness of pGGNs, and the similar performance was observed in the validation set.

CONCLUSION: The cLung-RADS® v2022 can effectively predict the invasiveness of pGGNs in real-world scenarios.

CRITICAL RELEVANCE STATEMENT: A complementary Lung-RADS® v2022 based on the Lung-RADS® v2022 and CT features can effectively predict the invasiveness of pulmonary pure ground-glass nodules and is applicable in clinical practice.

TRIAL REGISTRATION: Establishment and application of a multi-scale low-dose CT Lung cancer screening model based on modified lung-RADS1.1 and deep learning technology, 2022-KY-0137. Registered 24 January 2022. https://www.medicalresearch.org.cn/search/research/researchView?id=a97e67d8-1ee6-40fb-aab1-e6238dbd8f29 .

KEY POINTS: Lung-RADS® v2022 delayed lung cancer diagnosis for nodules appearing as pGGNs. Lung-RADS® v2022 showed lower accuracy and AUC than Lung-RADS 1.0. cLung-RADS® v2022 model effectively predicts the invasiveness of pulmonary pGGNs.

PMID:40121609 | DOI:10.1186/s13244-025-01937-3

BMC Musculoskelet Disord. 2025 Mar 22;26(1):287. doi: 10.1186/s12891-025-08538-7.

ABSTRACT

OBJECTIVES: Duchenne muscular dystrophy (DMD) and Becker muscular dystrophy (BMD) present similar symptoms in the early stage, complicating their differentiation. This study aims to develop a classification model using radiomic features from MRI T2-weighted Dixon sequences to increase the accuracy of distinguishing DMD and BMD in the early disease stage.

METHODS: We retrospectively analysed MRI data from 62 patients aged 36-60 months with muscular dystrophy, including 41 with DMD and 21 with BMD. Radiomic features were extracted from in-phase, opposed-phase, water, fat, and postprocessed fat fraction images. We employed a deep learning segmentation method to segment regions of interest automatically. Feature selection included the Mann‒Whitney U test for identifying significant features, Pearson correlation analysis to remove collinear features, and the LASSO regression method to select features with nonzero coefficients. These selected features were then used in various machine learning algorithms to construct the classification model, and their diagnostic performance was compared.

RESULTS: Our proposed radiomic and machine learning methods effectively distinguished early DMD and BMD. The machine learning models significantly outperformed the radiologists in terms of accuracy (81.2-90.6% compared with 69.4%), specificity (71.0-86.0% compared with 19.0%), and F1 score (85.2-92.6% compared with 80.5%), while maintaining relatively high sensitivity (85.6-95.0% compared with 95.1%).

CONCLUSION: Radiomics based on Dixon sequences combined with machine learning methods can effectively distinguish between DMD and BMD in the early stages, providing a new and effective tool for the early diagnosis of these muscular dystrophies.

CLINICAL TRIAL NUMBER: Not applicable.

PMID:40121488 | DOI:10.1186/s12891-025-08538-7

BMC Bioinformatics. 2025 Mar 22;26(1):88. doi: 10.1186/s12859-025-06101-8.

ABSTRACT

Binding proteins play a crucial role in biological systems by selectively interacting with specific molecules, such as DNA, RNA, or peptides, to regulate various cellular processes. Their ability to recognize and bind target molecules with high specificity makes them essential for signal transduction, transport, and enzymatic activity. Traditional experimental methods for identifying protein-binding peptides are costly and time-consuming. Current sequence-based approaches often struggle with accuracy, focusing too narrowly on proximal sequence features and ignoring structural data. This study presents Deep-ProBind, a powerful prediction model designed to classify protein binding sites by integrating sequence and structural information. The proposed model employs a transformer and evolutionary-based attention mechanism, i.e., Bidirectional Encoder Representations from Transformers (BERT) and Pseudo position specific scoring matrix -Discrete Wavelet Transform (PsePSSM -DWT) approach to encode peptides. The SHapley Additive exPlanations (SHAP) algorithm selects the optimal hybrid features, and a Deep Neural Network (DNN) is then used as the classification algorithm to predict protein-binding peptides. The performance of the proposed model was evaluated in comparison with traditional Machine Learning (ML) algorithms and existing models. Experimental results demonstrate that Deep-ProBind achieved 92.67% accuracy with tenfold cross-validation on benchmark datasets and 93.62% accuracy on independent samples. The Deep-ProBind outperforms existing models by 3.57% on training data and 1.52% on independent tests. These results demonstrate Deep-ProBind's reliability and effectiveness, making it a valuable tool for researchers and a potential resource in pharmacological studies, where peptide binding plays a critical role in therapeutic development.

PMID:40121399 | DOI:10.1186/s12859-025-06101-8

Sci Rep. 2025 Mar 22;15(1):9891. doi: 10.1038/s41598-025-94275-9.

ABSTRACT

In four-dimensional time series such as dynamic contrast-enhanced (DCE) MRI, motion between individual time steps due to the patient's breathing or movement leads to incorrect image analysis, e.g., when calculating perfusion. Image registration of the volumes of the individual time steps is necessary to improve the accuracy of the subsequent image analysis. Both groupwise and multiresolution registration methods have shown great potential for medical image registration. To combine the advantages of groupwise and multiresolution registration, we proposed a groupwise multiresolution network for deformable medical image registration. We applied our proposed method to the registration of DCE-MR images for the assessment of lung perfusion in patients with congenital diaphragmatic hernia. The networks were trained unsupervised with Mutual Information and Gradient L2 loss. We compared the groupwise networks with a pairwise deformable registration network and a published groupwise network as benchmarks and the classical registration method SimpleElastix as baseline using four-dimensional DCE-MR scans of patients after congenital diaphragmatic hernia repair. Experimental results showed that our groupwise network yields results with high spatial alignment (SSIM up to 0.953 ± 0.025 or 0.936 ± 0.028 respectively), medically plausible transformation with low image folding (|J| ≤ 0: 0.0 ± 0.0%), and a low registration time of less than 10 seconds for a four-dimensional DCE-MR scan with 50 time steps. Furthermore, our results demonstrate that image registration with the proposed groupwise network enhances the accuracy of medical image analysis by leading to more homogeneous perfusion maps.

PMID:40121309 | DOI:10.1038/s41598-025-94275-9

NPJ Digit Med. 2025 Mar 22;8(1):174. doi: 10.1038/s41746-025-01560-y.

ABSTRACT

Building accurate prediction models and identifying predictive biomarkers for treatment response in Muscle-Invasive Bladder Cancer (MIBC) are essential for improving patient survival but remain challenging due to tumor heterogeneity, despite numerous related studies. To address this unmet need, we developed an interpretable Graph-based Multimodal Late Fusion (GMLF) deep learning framework. Integrating histopathology and cell type data from standard H&E images with gene expression profiles derived from RNA sequencing from the SWOG S1314-COXEN clinical trial (ClinicalTrials.gov NCT02177695 2014-06-25), GMLF uncovered new histopathological, cellular, and molecular determinants of response to neoadjuvant chemotherapy. Specifically, we identified key gene signatures that drive the predictive power of our model, including alterations in TP63, CCL5, and DCN. Our discovery can optimize treatment strategies for patients with MIBC, e.g., improving clinical outcomes, avoiding unnecessary treatment, and ultimately, bladder preservation. Additionally, our approach could be used to uncover predictors for other cancers.

PMID:40121304 | DOI:10.1038/s41746-025-01560-y

Sci Rep. 2025 Mar 22;15(1):9972. doi: 10.1038/s41598-025-94464-6.

ABSTRACT

Recent data-driven deep learning methods for image reflection removal have made impressive progress, promoting the quality of photo capturing and scene understanding. Due to the massive consumption of computational complexity and memory usage, the performance of these methods degrades significantly while dealing with high-resolution images. Besides, most existing methods for reflection removal can only remove reflection patterns by downsampling the input image into a much lower resolution, resulting in the loss of plentiful information. In this paper, we propose a novel transformer-based framework for high-resolution image reflection removal, termed as the Laplacian pyramid-based component-aware transformer (LapCAT). LapCAT leverages a Laplacian pyramid network to remove high-frequency reflection patterns and reconstruct the high-resolution background image guided by the clean low-frequency background components. Guided by the reflection mask through pixel-wise contrastive learning, LapCAT designs a component-separable transformer block (CSTB) which removes reflection patterns from the background constituents through a reflection-aware multi-head self-attention mechanism. Extensive experiments on several benchmark datasets for reflection removal demonstrate the superiority of our LapCAT, especially the excellent performance and high efficiency in removing reflection from high-resolution images than state-of-the-art methods.

PMID:40121298 | DOI:10.1038/s41598-025-94464-6

Sci Rep. 2025 Mar 22;15(1):9924. doi: 10.1038/s41598-025-94437-9.

ABSTRACT

Medical image segmentation represents a pivotal and intricate procedure in the domain of medical image processing and analysis. With the progression of artificial intelligence in recent years, the utilization of deep learning techniques for medical image segmentation has witnessed escalating popularity. Nevertheless, the intricate nature of medical image poses challenges on the segmentation of diminutive targets is still in its early stages. Current networks encounter difficulties in addressing the segmentation of exceedingly small targets, especially when the number of training samples is limited. To overcome this constraint, we have implemented a proficient strategy to enhance lesion images containing small targets and constrained samples. We introduce a segmentation framework termed STS-Net, specifically designed for small target segmentation. This framework leverages the established capacity of convolutional neural networks to acquire effective image representations. The proposed STS-Net network adopts a ResNeXt50-32x4d architecture as the encoder, integrating attention mechanisms during the encoding phase to amplify the feature representation capabilities of the network. We evaluated the proposed network on four publicly available datasets. Experimental results underscore the superiority of our approach in the domain of medical image segmentation, particularly for small target segmentation. The codes are available at https://github.com/zlxokok/STSNet .

PMID:40121297 | DOI:10.1038/s41598-025-94437-9

Sci Rep. 2025 Mar 22;15(1):9905. doi: 10.1038/s41598-025-92872-2.

ABSTRACT

Predicting recurrence among early-stage hormone receptor-positive human epidermal growth factor receptor-negative breast cancer (HR+/HER2- BC) is crucial for guiding adjuvant therapy. However, studies are limited for patients with low recurrence risk. HR+/HER2- early-stage (T1-2N0-1) invasive BC patients who received definitive surgery and followed by endocrine therapy from four independent medical centers were included in this retrospective study. Patients from center 1 were used as derivation cohort, while those from other centers were combined as an external test cohort. A deep learning prognostic model, HERPAI, was developed based on Transformer to predict risk of invasive disease-free survival (iDFS) utilizing clinical and pathological predictors. The model performance was evaluated using C-index for the overall population and subgroups. Threshold for selecting 5-year recurrence risk > 10% was determined. Hazard ratio (HR) was estimated between risk groups for iDFS. A total of 6340 patients were included, of whom 5424 were assigned to the derivation cohort (training and validation [N = 4882] and internal test cohort [N = 542]), while 916 patients were utilized as external cohort. HERPAI yielded a C-index of 0.73 (95% CI 0.65-0.81), 0.73 (95% CI 0.62-0.85), and 0.68 (95% CI 0.60-0.77), in the validation, internal, and external test cohort, respectively. Consistent performances were observed for pre-specified subgroups. High-risk patients were associated with an increased risk of recurrence for validation (HR, 2.56 [95% CI 1.25-5.22], P = 0.01), internal test (HR, 2.52 [95% CI 0.97-6.57], P = 0.06) and external test (HR, 1.94 [95% CI 1.00-3.74], P = 0.049) cohort, respectively. HERPAI was a promising tool for selecting vulnerable early-stage HR+/HER2- BC patients who were at high-risk of recurrence. It could facilitate the prioritization of patients who may benefit more from escalating adjuvant treatment.

PMID:40121273 | DOI:10.1038/s41598-025-92872-2

Sci Rep. 2025 Mar 22;15(1):9914. doi: 10.1038/s41598-025-94267-9.

ABSTRACT

Accurate brain tumor segmentation is critical for clinical diagnosis and treatment planning. This study proposes an advanced segmentation framework that combines Multiscale Attention U-Net with the EfficientNetB4 encoder to enhance segmentation performance. Unlike conventional U-Net-based architectures, the proposed model leverages EfficientNetB4's compound scaling to optimize feature extraction at multiple resolutions while maintaining low computational overhead. Additionally, the Multi-Scale Attention Mechanism (utilizing [Formula: see text], and [Formula: see text] kernels) enhances feature representation by capturing tumor boundaries across different scales, addressing limitations of existing CNN-based segmentation methods. Our approach effectively suppresses irrelevant regions and enhances tumor localization through attention-enhanced skip connections and residual attention blocks. Extensive experiments were conducted on the publicly available Figshare brain tumor dataset, comparing different EfficientNet variants to determine the optimal architecture. EfficientNetB4 demonstrated superior performance, achieving an Accuracy of 99.79%, MCR of 0.21%, Dice Coefficient of 0.9339, and an Intersection over Union (IoU) of 0.8795, outperforming other variants in accuracy and computational efficiency. The training process was analyzed using key metrics, including Dice Coefficient, dice loss, precision, recall, specificity, and IoU, showing stable convergence and generalization. Additionally, the proposed method was evaluated against state-of-the-art approaches, surpassing them in all critical metrics, including accuracy, IoU, Dice Coefficient, precision, recall, specificity, and mean IoU. This study demonstrates the effectiveness of the proposed method for robust and efficient segmentation of brain tumors, positioning it as a valuable tool for clinical and research applications.

PMID:40121246 | DOI:10.1038/s41598-025-94267-9