Med Phys. 2025 Mar 18. doi: 10.1002/mp.17757. Online ahead of print.

ABSTRACT

BACKGROUND: Unsupervised domain adaptation (UDA) seeks to mitigate the performance degradation of deep neural networks when applied to new, unlabeled domains by leveraging knowledge from source domains. In medical image segmentation, prevailing UDA techniques often utilize adversarial learning to address domain shifts for cross-modality adaptation. Current research on adversarial learning tends to adopt increasingly complex models and loss functions, making the training process highly intricate and less stable/robust. Furthermore, most methods primarily focused on segmentation accuracy while neglecting the associated confidence levels and uncertainties.

PURPOSE: To develop a simple yet effective UDA method based on histogram matching-enhanced adversarial learning (HMeAL-UDA), and provide comprehensive uncertainty estimations of the model predictions.

METHODS: Aiming to bridge the domain gap while reducing the model complexity, we developed a novel adversarial learning approach to align multi-modality features. The method, termed HMeAL-UDA, integrates a plug-and-play histogram matching strategy to mitigate domain-specific image style biases across modalities. We employed adversarial learning to constrain the model in the prediction space, enabling it to focus on domain-invariant features during segmentation. Moreover, we quantified the model's prediction confidence using Monte Carlo (MC) dropouts to assess two voxel-level uncertainty estimates of the segmentation results, which were subsequently aggregated into a volume-level uncertainty score, providing an overall measure of the model's reliability. The proposed method was evaluated on three public datasets (Combined Healthy Abdominal Organ Segmentation [CHAOS], Beyond the Cranial Vault [BTCV], and Abdominal Multi-Organ Segmentation Challenge [AMOS]) and one in-house clinical dataset (UTSW). We used 30 MRI scans (20 from the CHAOS dataset and 10 from the in-house dataset) and 30 CT scans from the BTCV dataset for UDA-based, cross-modality liver segmentation. Additionally, 240 CT scans and 60 MRI scans from the AMOS dataset were utilized for cross-modality multi-organ segmentation. The training and testing sets for each modality were split with ratios of approximately 4:1-3:1.

RESULTS: Extensive experiments on cross-modality medical image segmentation demonstrated the superiority of HMeAL-UDA over two state-of-the-art approaches. HMeAL-UDA achieved a mean (± s.d.) Dice similarity coefficient (DSC) of 91.34% ± 1.23% and an HD95 of 6.18 ± 2.93 mm for cross-modality (from CT to MRI) adaptation of abdominal multi-organ segmentation, and a DSC of 87.13% ± 3.67% with an HD95 of 2.48 ± 1.56 mm for segmentation adaptation in the opposite direction (MRI to CT). The results are approaching or even outperforming those of supervised methods trained with "ground-truth" labels in the target domain. In addition, we provide a comprehensive assessment of the model's uncertainty, which can help with the understanding of segmentation reliability to guide clinical decisions.

CONCLUSION: HMeAL-UDA provides a powerful segmentation tool to address cross-modality domain shifts, with the potential to generalize to other deep learning applications in medical imaging.

PMID:40102198 | DOI:10.1002/mp.17757

Med Phys. 2025 Mar 18. doi: 10.1002/mp.17764. Online ahead of print.

ABSTRACT

BACKGROUND: To minimize radiation exposure while obtaining high-quality Positron Emission Tomography (PET) images, various methods have been developed to derive standard-count PET (SPET) images from low-count PET (LPET) images. Although deep learning methods have enhanced LPET images, they rarely utilize the rich complementary information from MR images. Even when MR images are used, these methods typically employ early, intermediate, or late fusion strategies to merge features from different CNN streams, failing to fully exploit the complementary properties of multimodal fusion.

PURPOSE: In this study, we introduce a novel multimodal feature-guided diffusion model, termed MFG-Diff, designed for the denoising of LPET images with the full utilization of MRI.

METHODS: MFG-Diff replaces random Gaussian noise with LPET images and introduces a novel degradation operator to simulate the physical degradation processes of PET imaging. Besides, it uses a novel cross-modal guided restoration network to fully exploit the modality-specific features provided by the LPET and MR images and utilizes a multimodal feature fusion module employing cross-attention mechanisms and positional encoding at multiple feature levels for better feature fusion.

RESULTS: Under four counts (2.5%, 5.0%, 10%, and 25%), the images generated by our proposed network showed superior performance compared to those produced by other networks in both qualitative and quantitative evaluations, as well as in statistical analysis. In particular, the peak-signal-to-noise ratio of the generated PET images improved by more than 20% under a 2.5% count, the structural similarity index improved by more than 16%, and the root mean square error reduced by nearly 50%. On the other hand, our generated PET images had significant correlation (Pearson correlation coefficient, 0.9924), consistency, and excellent quantitative evaluation results with the SPET images.

CONCLUSIONS: The proposed method outperformed existing state-of-the-art LPET denoising models and can be used to generate highly correlated and consistent SPET images obtained from LPET images.

PMID:40102174 | DOI:10.1002/mp.17764

ISA Trans. 2025 Mar 11:S0019-0578(25)00145-4. doi: 10.1016/j.isatra.2025.03.004. Online ahead of print.

ABSTRACT

Deep learning has significantly advanced bearing fault diagnosis. Traditional models rely on the assumption of independent and identically distributed, which is frequently violated due to variations in rotational speeds and loads during bearing fault diagnosis. The fault diagnosis of the bearing based on representation learning lacks the consideration of spectrum knowledge and representation diversity under multiple working conditions. Therefore, this study presents a domain-invariant representation learning strategy (DIRLs) for diagnosing bearing faults across differing working conditions. DIRLs, by leveraging envelope spectrum knowledge distillation, captures the Fourier characteristics as domain-invariant features and secures robust health state representations by aligning high-order statistics of the samples under different working conditions. Moreover, an innovative loss function, which maximizes the two-paradigm metric of the health state representation, is designed to enrich representation diversity. Experimental results demonstrate an average AUC improvement of 28.6 % on the Paderborn-bearing dataset and an overall diagnostic accuracy of 88.7 % on a private bearing dataset, validating the effectiveness of the proposed method.

PMID:40102111 | DOI:10.1016/j.isatra.2025.03.004

Z Med Phys. 2025 Mar 17:S0939-3889(25)00034-0. doi: 10.1016/j.zemedi.2025.02.004. Online ahead of print.

ABSTRACT

This work demonstrated the development of a fast, deep-learning framework (DeepSMCP) to mitigate noise in Monte Carlo dose distributions (MC-DDs) of photon treatment plans with high statistical uncertainty (SU) and its integration into the Swiss Monte Carlo Plan (SMCP). To this end, a two-channel input (MC-DD and computed tomography (CT) scan) 3D U-net was trained, validated and tested (80%/10%/10%) on high/low-SU MC-DD-pairs of 106 clinically-motivated VMAT arcs for 29 available CTs, augmented to 3074 pairs. The model was integrated into SMCP to enable a "one-click" workflow of calculating and denoising MC-DDs of high SU to obtain MC-DDs of low SU. The model accuracy was evaluated on the test set using Gamma passing rate (2% global, 2 mm, 10% threshold) comparing denoised and low-SU MC-DD. Calculation time for the whole workflow was recorded. Denoised MC-DDs match low-SU MC-DDs with average (standard deviation) Gamma passing rate of 82.9% (4.7%). Additional application of DeepSMCP to 12 unseen clinically-motivated cases of different treatment sites, including treatment sites not present during training, resulted in an average Gamma passing rate of 91.0%. Denoised DDs were obtained on average in 35.1 s, a 340-fold efficiency gain compared to low-SU MC-DD calculation. DeepSMCP presented a first seamlessly integrated promising denoising framework for MC-DDs.

PMID:40102103 | DOI:10.1016/j.zemedi.2025.02.004

J Chromatogr A. 2025 Mar 13;1748:465874. doi: 10.1016/j.chroma.2025.465874. Online ahead of print.

ABSTRACT

INTRODUCTION: High-Performance Liquid Chromatography (HPLC) is widely used for its high sensitivity, stability, and accuracy. Nonetheless, it often involves lengthy analysis times and considerable solvent consumption, especially when dealing with complex systems and quality control, posing challenges to green and eco-friendly analytical practices.

OBJECTIVE: This study proposes a compressed fingerprint chromatogram analysis technique that combines a one-dimensional convolutional neural network (1D-CNN) with HPLC, aiming to improve the analytical efficiency of various compounds in complex systems while reducing the use of organic solvents.

MATERIALS AND METHODS: The natural product Artemisiae argyi Folium (AAF) was selected as the experimental subject. Firstly, HPLC fingerprints of AAF were developed based on conventional programs. Next, a compressed fingerprint was obtained without losing compound information. Finally, a 1D-CNN deep learning model was used to analyze and identify the compressed chromatograms, enabling quantitative analysis of 10 compounds in complex systems.

RESULTS: The results indicate that the 1D-CNN model can effectively extract features from complex data, reducing the analysis time for each sample by about 40 min. In addition, the consumption of mobile phase has significantly decreased by 78 % compared to before. Among the ten compounds to be analyzed, nine of them achieved good results, with the highest correlation coefficient reaching above 0.95, indicating that the model has strong explanatory power.

CONCLUSION: The proposed compressed fingerprint chromatograms recognition technique enhances the environmental sustainability and efficiency of traditional HPLC methods, offering valuable insights for future advancements in analytical methodologies and equipment development.

PMID:40101658 | DOI:10.1016/j.chroma.2025.465874

Comput Biol Med. 2025 Mar 17;189:109970. doi: 10.1016/j.compbiomed.2025.109970. Online ahead of print.

ABSTRACT

Digital twins (DTs) are advancing biotechnology by providing digital models for drug discovery, digital health applications, and biological assets, including microorganisms. However, the hypothesis posits that implementing micro- and nanoscale DTs, especially for biological entities like bacteria, presents substantial challenges. These challenges stem from the complexities of data extraction, transmission, and computation, along with the necessity for a specialized Internet of Things (IoT) infrastructure. To address these challenges, this article proposes a novel framework that leverages bio-network technologies, including the Internet of Bio-Nano Things (IoBNT), and decentralized deep learning algorithms such as federated learning (FL) and convolutional neural networks (CNN). The methodology involves using CNNs for robust pattern recognition and FL to reduce bandwidth consumption while enhancing security. IoBNT devices are utilized for precise microscopic data acquisition and transmission, which ensures minimal error rates. The results demonstrate a multi-class classification accuracy of 98.7% across 33 bacteria categories, achieving over 99% bandwidth savings. Additionally, IoBNT integration reduces biological data transfer errors by up to 98%, even under worst-case conditions. This framework is further supported by an adaptable, user-friendly dashboard, expanding its applicability across pharmaceutical and biotechnology industries.

PMID:40101583 | DOI:10.1016/j.compbiomed.2025.109970

Ultrasonics. 2025 Mar 9;151:107631. doi: 10.1016/j.ultras.2025.107631. Online ahead of print.

ABSTRACT

Pipe wall loss assessment is crucial in oil and gas transportation. Ultrasonic guided wave is an effective technology to detect pipe defects. However, accurately inverting weak-feature defects under limited view conditions remains challenging due to constraints in transducer arrangements and inconsistent signal characteristics. This paper proposes a stepwise inversion method based on feature compensation and network reconstruction through deep learning, combined with high-order helical guided waves to expand the imaging view and achieve high-resolution imaging of pipe defects. A forward model was established using the finite difference method, with the two-dimensional Pearson correlation coefficient and maximum wall loss estimation accuracy defined as imaging metrics to evaluate and compare the method. Among 50 randomly selected defect samples in the test set, the inversion model achieved a correlation coefficient of 0.9669 and a maximum wall loss estimation accuracy of 96.65 %. Additionally, Gaussian noise was introduced to assess imaging robustness under pure signal, 5 dB, and 3 dB conditions. Laboratory experiments validated the practical feasibility of the proposed method. This approach is generalizable and holds significant potential for nondestructive testing in cylindrical waveguide structures represented by pipes.

PMID:40101471 | DOI:10.1016/j.ultras.2025.107631

Food Chem. 2025 Mar 12;479:143850. doi: 10.1016/j.foodchem.2025.143850. Online ahead of print.

ABSTRACT

The quality of chili peppers is closely related to their variety and geographical origin. The market often substitutes high-quality chili peppers with inferior ones, and cross-contamination occurs during processing. The existing methods cannot quickly and conveniently distinguish between different chili varieties or origins, which require expensive experimental equipment and professional skills. Techniques such as energy-dispersive X-ray fluorescence and inductively coupled plasma spectroscopy have been used for chili pepper classification and origin tracing, but these methods are either costly or destructive. To address the challenges of accurately identifying chili pepper varieties and origin tracing of chili peppers, this paper presents a sensor-aware convolutional network (SACNet) integrated with an electronic nose (e-nose) for accurate variety classification and origin traceability of chili peppers. The e-nose system collects gas samples from various chili peppers. We introduce a sensor attention module that adaptively focuses on the importance of each sensor in gathering gas information. Additionally, we introduce a local sensing and wide-area sensing structure to specifically capture gas information features, enabling high-precision identification of chili pepper gases. In comparative experiments with other networks, SACNet demonstrated excellent performance in both variety classification and origin traceability, and it showed significant advantages in terms of parameter quantity. Specifically, SACNet achieved 98.56 % accuracy in variety classification with Dataset A, 97.43 % accuracy in origin traceability with Dataset B, and 99.31 % accuracy with Dataset C. In summary, the combination of SACNet and an e-nose provides an effective strategy for identifying the varieties and origins of chili peppers.

PMID:40101378 | DOI:10.1016/j.foodchem.2025.143850

Med Image Anal. 2025 Mar 12;102:103542. doi: 10.1016/j.media.2025.103542. Online ahead of print.

ABSTRACT

Histopathological image classification using deep learning is crucial for accurate and efficient cancer diagnosis. However, annotating a large amount of histopathological images for training is costly and time-consuming, leading to a scarcity of available labeled data for training deep neural networks. To reduce human efforts and improve efficiency for annotation, we propose a Unified Semi-supervised Active Learning framework (UniSAL) that effectively selects informative and representative samples for annotation. First, unlike most existing active learning methods that only train from labeled samples in each round, dual-view high-confidence pseudo training is proposed to utilize both labeled and unlabeled images to train a model for selecting query samples, where two networks operating on different augmented versions of an input image provide diverse pseudo labels for each other, and pseudo label-guided class-wise contrastive learning is introduced to obtain better feature representations for effective sample selection. Second, based on the trained model at each round, we design novel uncertain and representative sample selection strategy. It contains a Disagreement-aware Uncertainty Selector (DUS) to select informative uncertain samples with inconsistent predictions between the two networks, and a Compact Selector (CS) to remove redundancy of selected samples. We extensively evaluate our method on three public pathological image classification datasets, i.e., CRC5000, Chaoyang and CRC100K datasets, and the results demonstrate that our UniSAL significantly surpasses several state-of-the-art active learning methods, and reduces the annotation cost to around 10% to achieve a performance comparable to full annotation. Code is available at https://github.com/HiLab-git/UniSAL.

PMID:40101375 | DOI:10.1016/j.media.2025.103542

Phys Med Biol. 2025 Mar 18. doi: 10.1088/1361-6560/adc236. Online ahead of print.

ABSTRACT

&#xD;To evaluate the impact of beam mask implementation and data aggregation on artificial intelligence-based dose prediction accuracy in proton therapy, with a focus on scenarios involving limited or highly heterogeneous datasets.&#xD;Approach:&#xD;In this study, 541 prostate and 632 head and neck (H&N) proton therapy plans were used to train and evaluate convolutional neural networks designed for the task of dose prediction. Datasets were grouped by anatomical site and beam configuration to assess the impact of beam masks-graphical depictions of radiation paths-as a model input. We also evaluated the effect of combining datasets. Model performance was measured using dose-volume histograms (DVH) scores, mean absolute error, mean absolute percent error, Dice similarity coefficients (DSC), and gamma passing rates.&#xD;Main results:&#xD;DSC analysis revealed that the inclusion of beam masks improved dose prediction accuracy, particularly in low-dose regions and for datasets with diverse beam configurations. Data aggregation alone produced mixed results, with improvements in high-dose regions but potential degradation in low-dose areas. Notably, combining beam masks and data aggregation yielded the best overall performance, effectively leveraging the strengths of both strategies. Additionally, the magnitude of the improvements was larger for datasets with greater heterogeneity, with the combined approach increasing the DSC score by as much as 0.2 for a subgroup of H&N cases characterized by small size and heterogeneity in beam arrangement. DVH scores reflected these benefits, showing statistically significant improvements (p < 0.05) for the more heterogeneous H&N datasets.&#xD;Significance:&#xD;Artificial intelligence-based dose prediction models incorporating beam masks and data aggregation significantly improve accuracy in proton therapy planning, especially for complex cases. This technique could accelerate the planning process, enabling more efficient and effective cancer treatment strategies.&#xD.

PMID:40101365 | DOI:10.1088/1361-6560/adc236

J Med Internet Res. 2025 Mar 18;27:e63962. doi: 10.2196/63962.

ABSTRACT

BACKGROUND: Monitoring the emotional states of patients with psychiatric problems has always been challenging due to the noncontinuous nature of clinical assessments, the effect of the health care environment, and the inherent subjectivity of evaluation instruments. However, mental states in psychiatric disorders exhibit substantial variability over time, making real-time monitoring crucial for preventing risky situations and ensuring appropriate treatment.

OBJECTIVE: This study aimed to leverage new technologies and deep learning techniques to enable more objective, real-time monitoring of patients. This was achieved by passively monitoring variables such as step count, patient location, and sleep patterns using mobile devices. We aimed to predict patient self-reports and detect sudden variations in their emotional valence, identifying situations that may require clinical intervention.

METHODS: Data for this project were collected using the Evidence-Based Behavior (eB2) app, which records both passive and self-reported variables daily. Passive data refer to behavioral information gathered via the eB2 app through sensors embedded in mobile devices and wearables. These data were obtained from studies conducted in collaboration with hospitals and clinics that used eB2. We used hidden Markov models (HMMs) to address missing data and transformer deep neural networks for time-series forecasting. Finally, classification algorithms were applied to predict several variables, including emotional state and responses to the Patient Health Questionnaire-9.

RESULTS: Through real-time patient monitoring, we demonstrated the ability to accurately predict patients' emotional states and anticipate changes over time. Specifically, our approach achieved high accuracy (0.93) and a receiver operating characteristic (ROC) area under the curve (AUC) of 0.98 for emotional valence classification. For predicting emotional state changes 1 day in advance, we obtained an ROC AUC of 0.87. Furthermore, we demonstrated the feasibility of forecasting responses to the Patient Health Questionnaire-9, with particularly strong performance for certain questions. For example, in question 9, related to suicidal ideation, our model achieved an accuracy of 0.9 and an ROC AUC of 0.77 for predicting the next day's response. Moreover, we illustrated the enhanced stability of multivariate time-series forecasting when HMM preprocessing was combined with a transformer model, as opposed to other time-series forecasting methods, such as recurrent neural networks or long short-term memory cells.

CONCLUSIONS: The stability of multivariate time-series forecasting improved when HMM preprocessing was combined with a transformer model, as opposed to other time-series forecasting methods (eg, recurrent neural network and long short-term memory), leveraging the attention mechanisms to capture longer time dependencies and gain interpretability. We showed the potential to assess the emotional state of a patient and the scores of psychiatric questionnaires from passive variables in advance. This allows real-time monitoring of patients and hence better risk detection and treatment adjustment.

PMID:40101216 | DOI:10.2196/63962

JMIR Med Educ. 2025 Mar 18;11:e55709. doi: 10.2196/55709.

ABSTRACT

BACKGROUND: Health care practitioners use clinical decision support systems (CDSS) as an aid in the crucial task of clinical reasoning and decision-making. Traditional CDSS are online repositories (ORs) and clinical practice guidelines (CPG). Recently, large language models (LLMs) such as ChatGPT have emerged as potential alternatives. They have proven to be powerful, innovative tools, yet they are not devoid of worrisome risks.

OBJECTIVE: This study aims to explore how medical students perform in an evaluated clinical case through the use of different CDSS tools.

METHODS: The authors randomly divided medical students into 3 groups, CPG, n=6 (38%); OR, n=5 (31%); and ChatGPT, n=5 (31%); and assigned each group a different type of CDSS for guidance in answering prespecified questions, assessing how students' speed and ability at resolving the same clinical case varied accordingly. External reviewers evaluated all answers based on accuracy and completeness metrics (score: 1-5). The authors analyzed and categorized group scores according to the skill investigated: differential diagnosis, diagnostic workup, and clinical decision-making.

RESULTS: Answering time showed a trend for the ChatGPT group to be the fastest. The mean scores for completeness were as follows: CPG 4.0, OR 3.7, and ChatGPT 3.8 (P=.49). The mean scores for accuracy were as follows: CPG 4.0, OR 3.3, and ChatGPT 3.7 (P=.02). Aggregating scores according to the 3 students' skill domains, trends in differences among the groups emerge more clearly, with the CPG group that performed best in nearly all domains and maintained almost perfect alignment between its completeness and accuracy.

CONCLUSIONS: This hands-on session provided valuable insights into the potential perks and associated pitfalls of LLMs in medical education and practice. It suggested the critical need to include teachings in medical degree courses on how to properly take advantage of LLMs, as the potential for misuse is evident and real.

PMID:40101183 | DOI:10.2196/55709

PLoS One. 2025 Mar 18;20(3):e0319679. doi: 10.1371/journal.pone.0319679. eCollection 2025.

ABSTRACT

The Stability of the economy is always a great challenge across the world, especially in under developed countries. Many researchers have contributed to forecasting the Stock Market and controlling the situation to ensure economic stability over the past several decades. For this purpose, many researchers have built various models and gained benefits. This journey continues to date and will persist for the betterment of the stock market. This study is also a part of this journey, where four learning-based models are tailored for stock price prediction. Daily business data from the Karachi Stock Exchange (100 Index), covering from February 22, 2008 to February 23, 2021, is used for training and testing these models. This paper presenting four deep learning models with different architectures, namely the Artificial Neural Network model, the Recurrent Neural Network with Attention model, the Long Short-Term Memory Network with Attention model, and the Gated Recurrent Unit with Attention model. The Long Short-Term Memory with attention model was found to be the top-performing technique for accurately predicting stock exchange prices. During the Training, Validation and Testing Sessions, we observed the R-Squared values of the proposed model to be 0.9996, 0.9980 and 0.9921, respectively, making it the best-performing model among those mentioned above.

PMID:40100866 | DOI:10.1371/journal.pone.0319679

PLoS One. 2025 Mar 18;20(3):e0317863. doi: 10.1371/journal.pone.0317863. eCollection 2025.

ABSTRACT

Efficient image retrieval from a variety of datasets is crucial in today's digital world. Visual properties are represented using primitive image signatures in Content Based Image Retrieval (CBIR). Feature vectors are employed to classify images into predefined categories. This research presents a unique feature identification technique based on suppression to locate interest points by computing productive sum of pixel derivatives by computing the differentials for corner scores. Scale space interpolation is applied to define interest points by combining color features from spatially ordered L2 normalized coefficients with shape and object information. Object based feature vectors are formed using high variance coefficients to reduce the complexity and are converted into bag-of-visual-words (BoVW) for effective retrieval and ranking. The presented method encompass feature vectors for information synthesis and improves the discriminating strength of the retrieval system by extracting deep image features including primitive, spatial, and overlayed using multilayer fusion of Convolutional Neural Networks(CNNs). Extensive experimentation is performed on standard image datasets benchmarks, including ALOT, Cifar-10, Corel-10k, Tropical Fruits, and Zubud. These datasets cover wide range of categories including shape, color, texture, spatial, and complicated objects. Experimental results demonstrate considerable improvements in precision and recall rates, average retrieval precision and recall, and mean average precision and recall rates across various image semantic groups within versatile datasets. The integration of traditional feature extraction methods fusion with multilevel CNN advances image sensing and retrieval systems, promising more accurate and efficient image retrieval solutions.

PMID:40100801 | DOI:10.1371/journal.pone.0317863

PLoS One. 2025 Mar 18;20(3):e0320337. doi: 10.1371/journal.pone.0320337. eCollection 2025.

NO ABSTRACT

PMID:40100785 | DOI:10.1371/journal.pone.0320337

IEEE Trans Neural Syst Rehabil Eng. 2025 Mar 18;PP. doi: 10.1109/TNSRE.2025.3552530. Online ahead of print.

ABSTRACT

Current surface electromyography (sEMG) methods for locomotion mode prediction face limitations in anticipatory capability due to computation delays and constrained window lengths typically below 500ms-a practice historically tied to stationarity requirements of handcrafted feature extraction. This study investigates whether end-to-end convolutional neural networks (CNNs) processing raw sEMG signals can overcome these constraints through extended window lengths (250ms to 1500 ms). We systematically evaluate six window lengths paired with three prediction horizons (model forecasts 50ms to 150ms ahead) in a continuous locomotion task involving eight modes and 16 transitions. The optimal configuration (1000ms window with 150ms horizon) achieved subject-average accuracies of 96.93% (steady states) and 97.50% (transient states), maintaining 95.03% and 85.53% respectively in real-time simulations. With a net averaged anticipation time of 147.9ms after 2.1ms computation latency, this approach demonstrates that windows covering 74% of the gait cycle can synergize with deep learning to balance the inherent trade-off between extracting richer information and maintaining system responsiveness to changes in activity.

PMID:40100693 | DOI:10.1109/TNSRE.2025.3552530

IEEE J Biomed Health Inform. 2025 Mar 18;PP. doi: 10.1109/JBHI.2025.3551720. Online ahead of print.

ABSTRACT

The intelligent analysis of Whole Slide Images (WSI) in digital pathology is critical for advancing precision medicine, particularly in oncology. However, the availability of WSI datasets is often limited by privacy regulations, which constrains the performance and generalizability of deep learning models. To address this challenge, this paper proposes an improved data augmentation method based on Deep Convolutional Generative Adversarial Network (DCGAN). Our approach leverages self-supervised pretraining with the CTransPath model to extract diverse and representationally rich WSI features, which guide the generation of high-quality synthetic images. We further enhance the model by introducing a least-squares adversarial loss and a frequency domain loss to improve pixel-level accuracy and structural fidelity, while incorporating residual blocks and skip connections to increase network depth, mitigate gradient vanishing, and improve training stability. Experimental results on the PatchCamelyon dataset demonstrate that our improved DCGAN achieves superior SSIM and FID scores compared to traditional models. The augmented datasets significantly enhance the performance of downstream classification tasks, improving accuracy, AUC, and F1 scores.

PMID:40100674 | DOI:10.1109/JBHI.2025.3551720

IEEE Trans Biomed Eng. 2025 Mar 18;PP. doi: 10.1109/TBME.2025.3550842. Online ahead of print.

ABSTRACT

OBJECTIVE: Predicting normative pediatric growth is crucial to identify developmental anomalies. While traditional statistical and computational methods have shown promising results predicting personalized development, they either rely on statistical assumptions that limit generalizability or require longitudinal datasets, which are scarce in children. Recent deep learning methods trained with cross-sectional dataset have shown potential to predict temporal changes but have only succeeded at predicting local intensity changes and can hardly model major anatomical changes that occur during childhood. We present a novel deep learning method for image synthesis that can be trained using only cross-sectional data to make personalized predictions of pediatric development.

METHODS: We designed a new generative adversarial network (GAN) with a novel Siamese cyclic encoder-decoder generator architecture and an identity preservation mechanism. Our design allows the encoder to learn age- and sex-independent identity-preserving representations of patient phenotypes from single images by leveraging the statistical distributions in the cross-sectional dataset. The decoder learns to synthesize personalized images from the encoded representations at any age.

RESULTS: Trained using only cross-sectional head CT images from 2,014 subjects (age 0-10 years), our model demonstrated state-of-the-art performance evaluated on an independent longitudinal dataset with images from 51 subjects.

CONCLUSION: Our method can predict pediatric development and synthesize temporal image sequences with state-of-the-art accuracy without requiring longitudinal images for training.

SIGNIFICANCE: Our method enables the personalized prediction of pediatric growth and longitudinal synthesis of clinical images, hence providing a patient-specific reference of normative development.

PMID:40100672 | DOI:10.1109/TBME.2025.3550842

IEEE Trans Neural Netw Learn Syst. 2025 Mar 18;PP. doi: 10.1109/TNNLS.2025.3540291. Online ahead of print.

ABSTRACT

Protein complex structural data are growing at an unprecedented pace, but its complexity and diversity pose significant challenges for protein function research. Although deep learning models have been widely used to capture the syntactic structure, word semantics, or semantic meanings of polypeptide and protein sequences, these models often overlook the complex contextual information of sequences. Here, we propose interpretable interaction deep learning (IIDL)-peptide-protein interaction (PepPI), a deep learning model designed to tackle these challenges using data-driven and interpretable pragmatic analysis to profile PepPIs. IIDL-PepPI constructs bidirectional attention modules to represent the contextual information of peptides and proteins, enabling pragmatic analysis. It then adopts a progressive transfer learning framework to simultaneously predict PepPIs and identify binding residues for specific interactions, providing a solution for multilevel in-depth profiling. We validate the performance and robustness of IIDL-PepPI in accurately predicting peptide-protein binary interactions and identifying binding residues compared with the state-of-the-art methods. We further demonstrate the capability of IIDL-PepPI in peptide virtual drug screening and binding affinity assessment, which is expected to advance artificial intelligence-based peptide drug discovery and protein function elucidation.

PMID:40100664 | DOI:10.1109/TNNLS.2025.3540291

IEEE Trans Pattern Anal Mach Intell. 2025 Mar 18;PP. doi: 10.1109/TPAMI.2025.3552484. Online ahead of print.

ABSTRACT

The divergence between labeled training data and unlabeled testing data is a significant challenge for recent deep learning models. Unsupervised domain adaptation (UDA) attempts to solve such problem. Recent works show that self-training is a powerful approach to UDA. However, existing methods have difficulty in balancing the scalability and performance. In this paper, we propose a hard-aware instance adaptive self-training framework for UDA on the task of semantic segmentation. To effectively improve the quality and diversity of pseudo-labels, we develop a novel pseudo-label generation strategy with an instance adaptive selector. We further enrich the hard class pseudo-labels with inter-image information through a skillfully designed hard-aware pseudo-label augmentation. Besides, we propose the region-adaptive regularization to smooth the pseudo-label region and sharpen the non-pseudo-label region. For the non-pseudo-label region, consistency constraint is also constructed to introduce stronger supervision signals during model optimization. Our method is so concise and efficient that it is easy to be generalized to other UDA methods. Experiments on GTA5 Cityscapes, SYNTHIA Cityscapes, and Cityscapes Oxford RobotCar demonstrate the superior performance of our approach compared with the state-of-the-art methods. Our codes are available at https://github.com/bupt-ai-cz/HIAST.

PMID:40100655 | DOI:10.1109/TPAMI.2025.3552484