Deep learning
Multifactor prediction model for stock market analysis based on deep learning techniques
Sci Rep. 2025 Feb 11;15(1):5121. doi: 10.1038/s41598-025-88734-6.
ABSTRACT
Stock market stability relies on the shares, investors, and stakeholders' participation and global commodity exchanges. In general, multiple factors influence the stock market stability to ensure profitable returns and commodity transactions. This article presents a contradictory-factor-based stability prediction model using the sigmoid deep learning paradigm. Sigmoid learning identifies the possible stabilizations of different influencing factors toward a profitable stock exchange. In this model, each influencing factor is mapped with the profit outcomes considering the live shares and their exchange value. The stability is predicted using sigmoid and non-sigmoid layers repeatedly until the maximum is reached. This stability is matched with the previous outcomes to predict the consecutive hours of stock market changes. Based on the actual changes and predicted ones, the sigmoid function is altered to accommodate the new range. The non-sigmoid layer remains unchanged in the new changes to improve the prediction precision. Based on the outcomes the deep learning's sigmoid layer is trained to identify abrupt changes in the stock market.
PMID:39934296 | DOI:10.1038/s41598-025-88734-6
Interpretable deep learning of single-cell and epigenetic data reveals novel molecular insights in aging
Sci Rep. 2025 Feb 11;15(1):5048. doi: 10.1038/s41598-025-89646-1.
ABSTRACT
Deep learning (DL) and explainable artificial intelligence (XAI) have emerged as powerful machine-learning tools to identify complex predictive data patterns in a spatial or temporal domain. Here, we consider the application of DL and XAI to large omic datasets, in order to study biological aging at the molecular level. We develop an advanced multi-view graph-level representation learning (MGRL) framework that integrates prior biological network information, to build molecular aging clocks at cell-type resolution, which we subsequently interpret using XAI. We apply this framework to one of the largest single-cell transcriptomic datasets encompassing over a million immune cells from 981 donors, revealing a ribosomal gene subnetwork, whose expression correlates with age independently of cell-type. Application of the same DL-XAI framework to DNA methylation data of sorted monocytes reveals an epigenetically deregulated inflammatory response pathway whose activity increases with age. We show that the ribosomal module and inflammatory pathways would not have been discovered had we used more standard machine-learning methods. In summary, the computational deep learning framework presented here illustrates how deep learning when combined with explainable AI tools, can reveal novel biological insights into the complex process of aging.
PMID:39934290 | DOI:10.1038/s41598-025-89646-1
A promising AI based super resolution image reconstruction technique for early diagnosis of skin cancer
Sci Rep. 2025 Feb 11;15(1):5084. doi: 10.1038/s41598-025-89693-8.
ABSTRACT
Skin cancer can be prevalent in people of any age group who are exposed to ultraviolet (UV) radiation. Among all other types, melanoma is a notable severe kind of skin cancer, which can be fatal. Melanoma is a malignant skin cancer arising from melanocytes, requiring early detection. Typically, skin lesions are classified either as benign or malignant. However, some lesions do exist that don't show clear cancer signs, making them suspicious. If unnoticed, these suspicious lesions develop into severe melanoma, requiring invasive treatments later on. These intermediate or suspicious skin lesions are completely curable if it is diagnosed at their early stages. To tackle this, few researchers intended to improve the image quality of the infected lesions obtained from the dermoscopy through image reconstruction techniques. Analyzing reconstructed super-resolution (SR) images allows early detection, fine feature extraction, and treatment plans. Despite advancements in machine learning, deep learning, and complex neural networks enhancing skin lesion image quality, a key challenge remains unresolved: how the intricate textures are obtained while performing significant up scaling in medical image reconstruction? Thus, an artificial intelligence (AI) based reconstruction algorithm is proposed to obtain the fine features from the intermediate skin lesion from dermoscopic images for early diagnosis. This serves as a non-invasive approach. In this research, a novel melanoma information improvised generative adversarial network (MELIIGAN) framework is proposed for the expedited diagnosis of intermediate skin lesions. Also, designed a stacked residual block that handles larger scaling factors and the reconstruction of fine-grained details. Finally, a hybrid loss function with a total variation (TV) regularization term switches to the Charbonnier loss function, a robust substitute for the mean square error loss function. The benchmark dataset results in a structural index similarity (SSIM) of 0.946 and a peak signal-to-noise ratio (PSNR) of 40.12 dB as the highest texture information, evidently compared to other state-of-the-art methods.
PMID:39934265 | DOI:10.1038/s41598-025-89693-8
RAE-Net: a multi-modal neural network based on feature fusion and evidential deep learning algorithm in predicting breast cancer subtypes on DCE-MRI
Biomed Phys Eng Express. 2025 Feb 11. doi: 10.1088/2057-1976/adb494. Online ahead of print.
ABSTRACT
Abstract

Objectives: Accurate identification of molecular subtypes in breast cancer is critical for personalized treatment. This study introduces a novel neural network model, RAE-Net, based on Multimodal Feature Fusion (MFF) and the Evidential Deep Learning Algorithm (EDLA) to improve breast cancer subtype prediction using dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI).

Methods: A dataset of 344 patients with histologically confirmed breast cancer was divided into training (n=200), validation (n=60), and testing (n=62) cohorts. RAE-Net, built on ResNet-50 with Multi-Head Attention (MHA) fusion and Multi-Layer Perceptron (MLP) mechanisms, combines radiomic and deep learning features for subtype prediction. The EDLA module adds uncertainty estimation to enhance classification reliability.

Results: RAE-Net with the MFF module achieved a mean accuracy of 0.83 and a Macro-F1 score of 0.78, outperforming traditional radiomics models (accuracy: 0.79, Macro-F1: 0.75) and standalone deep learning models (accuracy: 0.80, Macro-F1: 0.76). With an EDLA uncertainty threshold of 0.2, performance improved significantly, with accuracy reaching 0.97 and Macro-F1 at 0.92. RAE-Net also outperformed ResGANet, increasing accuracy by 0.5% and improving AUC. Compared to HIFUSE, RAE-Net reduced parameters and computational cost by 90%, with only a 5.7% increase in computation time.

Conclusions: RAE-Net integrates feature fusion and uncertainty estimation to predict breast cancer subtypes from DCE-MRI. The model achieves high accuracy while maintaining computational efficiency, demonstrating its potential for clinical use as a reliable and resource-efficient diagnostic tool.
PMID:39933196 | DOI:10.1088/2057-1976/adb494
Deep learning-assisted identification and localization of ductal carcinoma from bulk tissue in-silico models generated through polarized Monte Carlo simulations
Biomed Phys Eng Express. 2025 Feb 11. doi: 10.1088/2057-1976/adb495. Online ahead of print.
ABSTRACT
Despite significant progress in diagnosis and treatment, breast cancer remains a formidable health challenge, emphasizing the continuous need for research. This simulation study uses polarized Monte Carlo approach to identify and locate breast cancer. The tissue model Mueller matrix derived from polarized Monte Carlo simulations provides enhanced contrast for better comprehension of tissue structures. This study explicitly targets tumour regions found at the tissue surface, a possible scenario in thick tissue sections obtained after surgical removal of breast tissue lumps. We use a convolutional neural network for the identification and localization of tumours. Nine distinct spatial positions, defined relative to the point of illumination, allow the identification of the tumour even if it is outside the directly illuminated area. A system incorporating deep learning techniques automates processes and enables real-time diagnosis. This research paper aims to showcase the concurrent detection of the tumour's existence and position by utilizing a Convolutional Neural Network (CNN) implemented on depolarized index images derived from polarized Monte Carlo simulations. The classification accuracy achieved by the CNN model stands at 96%, showcasing its optimal performance. The model is also tested with images obtained from in-vitro tissue models, which yielded 100% classification accuracy on a selected subset of spatial positions.
PMID:39933195 | DOI:10.1088/2057-1976/adb495
Deep Learning-based Video-level View Classification of Two-dimensional Transthoracic Echocardiography
Biomed Phys Eng Express. 2025 Feb 11. doi: 10.1088/2057-1976/adb493. Online ahead of print.
ABSTRACT
In recent years, deep learning (DL)-based automatic view classification of 2D transthoracic echocardiography (TTE) has demonstrated strong performance, but has not fully addressed key clinical requirements such as view coverage, classification accuracy, inference delay, and the need for thorough exploration of performance in real-world clinical settings. We proposed a clinical requirement-driven DL framework, TTESlowFast, for accurate and efficient video-level TTE view classification. This framework is based on the SlowFast architecture and incorporates both a sampling balance strategy and a data augmentation strategy to address class imbalance and the limited availability of labeled TTE videos, respectively. TTESlowFast achieved an overall accuracy of 0.9881, precision of 0.9870, recall of 0.9867, and F1 score of 0.9867 on the test set. After field deployment, the model's overall accuracy, precision, recall, and F1 score for view classification were 0.9607, 0.9586, 0.9499, and 0.9530, respectively. The inference time for processing a single TTE video was (105.0 ± 50.1) ms on a desktop GPU (NVIDIA RTX 3060) and (186.0 ± 5.2) ms on an edge computing device (Jetson Orin Nano), which basically meets the clinical demand for immediate processing following image acquisition. The TTESlowFast framework proposed in this study demonstrates effective performance in TTE view classification with low inference delay, making it well-suited for various medical scenarios and showing significant potential for practical application.
PMID:39933194 | DOI:10.1088/2057-1976/adb493
A deep learning-enabled smart garment for accurate and versatile monitoring of sleep conditions in daily life
Proc Natl Acad Sci U S A. 2025 Feb 18;122(7):e2420498122. doi: 10.1073/pnas.2420498122. Epub 2025 Feb 11.
ABSTRACT
In wearable smart systems, continuous monitoring and accurate classification of different sleep-related conditions are critical for enhancing sleep quality and preventing sleep-related chronic conditions. However, the requirements for device-skin coupling quality in electrophysiological sleep monitoring systems hinder the comfort and reliability of night wearing. Here, we report a washable, skin-compatible smart garment sleep monitoring system that captures local skin strain signals under weak device-skin coupling conditions without positioning or skin preparation requirements. A printed textile-based strain sensor array responds to strain from 0.1 to 10% with a gauge factor as high as 100 and shows independence to extrinsic motion artifacts via strain-isolating printed pattern design. Through reversible starching treatment, ink penetration depth during direct printing on garments is controlled to achieve batch-to-batch performance variation <10%. Coupled with deep learning, explainable AI, and transfer learning data processing, the smart garment is capable of classifying six sleep states with an accuracy of 98.6%, maintaining excellent explainability (classification with low bias) and generalization (95% accuracy on new users with few-shot learning less than 15 samples per class) in practical applications, paving the way for next-generation daily sleep healthcare management.
PMID:39932995 | DOI:10.1073/pnas.2420498122
A Conditional Denoising VAE-based Framework for Antimicrobial Peptides Generation with Preserving Desirable Properties
Bioinformatics. 2025 Feb 11:btaf069. doi: 10.1093/bioinformatics/btaf069. Online ahead of print.
ABSTRACT
MOTIVATION: The widespread use of antibiotics has led to the emergence of resistant pathogens. Antimicrobial peptides (AMPs) combat bacterial infections by disrupting the integrity of cell membranes, making it challenging for bacteria to develop resistance. Consequently, AMPs offer a promising solution to addressing antibiotic resistance. However, the limited availability of natural AMPs cannot meet the growing demand. While deep learning technologies have advanced AMP generation, conventional models often lack stability and may introduce unforeseen side effects.
RESULTS: This study presents a novel denoising VAE-based model guided by desirable physicochemical properties for AMPs generation. The model integrates key features (e.g., molecular weight, isoelectric point, hydrophobicity, etc.), and employs position encoding along with a Transformer architecture to enhance generation accuracy. A customized loss function, combining reconstruction loss, KL divergence, and property preserving loss, ensures effective model training. Additionally, the model incorporates a denoising mechanism, enabling it to learn from perturbed inputs, thus maintaining performance under limited training data. Experimental results demonstrate that the proposed model can generate AMPs with desirable functional properties, offering a viable approach for AMP design and analysis, which ultimately contributes to the fight against antibiotic resistance.
AVAILABILITY AND IMPLEMENTATION: The data and source codes are available both in GitHub (https://github.com/David-WZhao/PPGC-DVAE) and Zenodo (DOI 10.5281/zenodo.14730711).
CONTACT AND SUPPLEMENTARY INFORMATION: wzzhao@ccnu.edu.cn, and Supplementary materials are available at Bioinformatics online.
PMID:39932977 | DOI:10.1093/bioinformatics/btaf069
A privacy-preserved horizontal federated learning for malignant glioma tumour detection using distributed data-silos
PLoS One. 2025 Feb 11;20(2):e0316543. doi: 10.1371/journal.pone.0316543. eCollection 2025.
ABSTRACT
Malignant glioma is the uncontrollable growth of cells in the spinal cord and brain that look similar to the normal glial cells. The most essential part of the nervous system is glial cells, which support the brain's functioning prominently. However, with the evolution of glioma, tumours form that invade healthy tissues in the brain, leading to neurological impairment, seizures, hormonal dysregulation, and venous thromboembolism. Medical tests, including medical resonance imaging (MRI), computed tomography (CT) scans, biopsy, and electroencephalograms are used for early detection of glioma. However, these tests are expensive and may cause irritation and allergic reactions due to ionizing radiation. The deep learning models are highly optimal for disease prediction, however, the challenge associated with it is the requirement for substantial memory and storage to amalgamate the patient's information at a centralized location. Additionally, it also has patient data-privacy concerns leading to anonymous information generalization, regulatory compliance issues, and data leakage challenges. Therefore, in the proposed work, a distributed and privacy-preserved horizontal federated learning-based malignant glioma disease detection model has been developed by employing 5 and 10 different clients' architectures in independent and identically distributed (IID) and non-IID distributions. Initially, for developing this model, the collection of the MRI scans of non-tumour and glioma tumours has been done, which are further pre-processed by performing data balancing and image resizing. The configuration and development of the pre-trained MobileNetV2 base model have been performed, which is then applied to the federated learning(FL) framework. The configurations of this model have been kept as 0.001, Adam, 32, 10, 10, FedAVG, and 10 for learning rate, optimizer, batch size, local epochs, global epochs, aggregation, and rounds, respectively. The proposed model has provided the most prominent accuracy with 5 clients' architecture as 99.76% and 99.71% for IID and non-IID distributions, respectively. These outcomes demonstrate that the model is highly optimized and generalizes the improved outcomes when compared to the state-of-the-art models.
PMID:39932966 | DOI:10.1371/journal.pone.0316543
Deep learning-based segmentation of the mandibular canals in cone beam computed tomography reaches human level performance
Dentomaxillofac Radiol. 2025 Feb 11:twae069. doi: 10.1093/dmfr/twae069. Online ahead of print.
ABSTRACT
OBJECTIVES: This study evaluated the accuracy and reliability of deep learning-based segmentation techniques for mandibular canal identification in CBCT data to provide a reliable and efficient support-tool for dental implant treatment planning.
METHODS: A dataset of 90 cone beam computed tomography (CBCT) scans was annotated as ground truth for mandibular canal segmentation. The dataset was split into training (n = 69), validation (n = 1), and testing (n = 20) subsets. A deep learning model based on a hierarchical convolutional neural network architecture was developed and trained. The model's performance was evaluated using Dice similarity coefficient (DSC), 95% Hausdorff distance (HD), and average symmetric surface distance (ASSD). Qualitative assessment was performed by two experienced dental imaging practitioners who evaluated the segmentation quality in terms of trust and safety on a 5-point Likert scale at three mandibular locations per side.
RESULTS: The trained model achieved a mean DSC of 0.77 ± 0.09, HD of 1.66 ± 0.86 mm, and ASSD of 0.31 ± 0.15 mm on the testing subset. Qualitative assessment showed no significant difference between the deep learning-based segmentations and ground truth in terms of trust and safety across all investigated locations (p > 0.05).
CONCLUSIONS: The proposed deep learning-based segmentation technique exhibits sufficient accuracy for the reliable identification of mandibular canals in CBCT scans. This automated approach could streamline the pre-operative planning process for dental implant placement, reducing the risk of neurovascular complications and enhancing patient safety.
PMID:39932925 | DOI:10.1093/dmfr/twae069
Enhancing PM2.5 prediction by mitigating annual data drift using wrapped loss and neural networks
PLoS One. 2025 Feb 11;20(2):e0314327. doi: 10.1371/journal.pone.0314327. eCollection 2025.
ABSTRACT
In many deep learning tasks, it is assumed that the data used in the training process is sampled from the same distribution. However, this may not be accurate for data collected from different contexts or during different periods. For instance, the temperatures in a city can vary from year to year due to various unclear reasons. In this paper, we utilized three distinct statistical techniques to analyze annual data drifting at various stations. These techniques calculate the P values for each station by comparing data from five years (2014-2018) to identify data drifting phenomena. To find out the data drifting scenario those statistical techniques and calculate the P value from those techniques to measure the data drifting in specific locations. From those statistical techniques, the highest drifting stations can be identified from the previous year's datasets To identify data drifting and highlight areas with significant drift, we utilized meteorological air quality and weather data in this study. We proposed two models that consider the characteristics of data drifting for PM2.5 prediction and compared them with various deep learning models, such as Long Short-Term Memory (LSTM) and its variants, for predictions from the next hour to the 64th hour. Our proposed models significantly outperform traditional neural networks. Additionally, we introduced a wrapped loss function incorporated into a model, resulting in more accurate results compared to those using the original loss function alone and prediction has been evaluated by RMSE, MAE and MAPE metrics. The proposed Front-loaded connection model(FLC) and Back-loaded connection model (BLC) solve the data drifting issue and the wrap loss function also help alleviate the data drifting problem with model training and works for the neural network models to achieve more accurate results. Eventually, the experimental results have shown that the proposed model performance enhanced from 24.1% -16%, 12%-8.3% respectively at 1h-24h, 32h-64h with compared to baselines BILSTM model, by 24.6% -11.8%, 10%-10.2% respectively at 1h-24h, 32h-64h compared to CNN model in hourly PM2.5 predictions.
PMID:39932913 | DOI:10.1371/journal.pone.0314327
Quantifying multilabeled brain cells in the whole prefrontal cortex reveals reduced inhibitory and a subtype of excitatory neuronal marker expression in serotonin transporter knockout rats
Cereb Cortex. 2025 Feb 5;35(2):bhae486. doi: 10.1093/cercor/bhae486.
ABSTRACT
The prefrontal cortex regulates emotions and is influenced by serotonin. Rodents lacking the serotonin transporter (5-HTT) show increased anxiety and changes in excitatory and inhibitory cell markers in the prefrontal cortex. However, these observations are constrained by limitations in brain representation and cell segmentation, as standard immunohistochemistry is inadequate to consider volume variations in regions of interest. We utilized the deep learning network of the StarDist method in combination with novel open-source methods for automated cell counts in a wide range of prefrontal cortex subregions. We found that 5-HTT knockout rats displayed increased anxiety and diminished relative numbers of subclass excitatory VGluT2+ and activated ΔFosB+ cells in the infralimbic and prelimbic cortices and of inhibitory GAD67+ cells in the prelimbic cortex. Anxiety levels and ΔFosB cell counts were positively correlated in wild-type, but not in knockout, rats. In conclusion, we present a novel method to quantify whole brain subregions of multilabeled cells in animal models and demonstrate reduced excitatory and inhibitory neuronal marker expression in prefrontal cortex subregions of 5-HTT knockout rats.
PMID:39932853 | DOI:10.1093/cercor/bhae486
Does Deep Learning Reconstruction Improve Ureteral Stone Detection and Subjective Image Quality in the CT Images of Patients with Metal Hardware?
J Endourol. 2025 Feb 11. doi: 10.1089/end.2024.0666. Online ahead of print.
ABSTRACT
Introduction: Diagnosing ureteral stones with low-dose CT in patients with metal hardware can be challenging because of image noise. The purpose of this study was to compare ureteral stone detection and image quality of low-dose and conventional CT scans with and without deep learning reconstruction (DLR) and metal artifact reduction (MAR) in the presence of metal hip prostheses. Methods: Ten urinary system combinations with 4 to 6 mm ureteral stones were implanted into a cadaver with bilateral hip prostheses. Each set was scanned under two different radiation doses (conventional dose [CD] = 115 mAs and ultra-low dose [ULD] = 6.0 mAs). Two scans were obtained for each dose as follows: one with and another without DLR and MAR. Two blinded radiologists ranked each image in terms of artifact, image noise, image sharpness, overall quality, and diagnostic confidence. Stone detection accuracy at each setting was calculated. Results: ULD with DLR and MAR improved subjective image quality in all five domains (p < 0.05) compared with ULD. In addition, the subjective image quality for ULD with DLR and MAR was greater than the subjective image quality for CD in all five domains (p < 0.05). Stone detection accuracy of ULD improved with the application of DLR and MAR (p < 0.05). Stone detection accuracy of ULD with DLR and MAR was similar to CD (p > 0.25). Conclusions: DLR with MAR may allow the application of low-dose CT protocols in patients with hip prostheses. Application of DLR and MAR to ULD provided a stone detection accuracy comparable with CD, reduced radiation exposure by 94.8%, and improved subjective image quality.
PMID:39932744 | DOI:10.1089/end.2024.0666
Diffusion-driven multi-modality medical image fusion
Med Biol Eng Comput. 2025 Feb 11. doi: 10.1007/s11517-025-03300-6. Online ahead of print.
ABSTRACT
Multi-modality medical image fusion (MMIF) technology utilizes the complementarity of different modalities to provide more comprehensive diagnostic insights for clinical practice. Existing deep learning-based methods often focus on extracting the primary information from individual modalities while ignoring the correlation of information distribution across different modalities, which leads to insufficient fusion of image details and color information. To address this problem, a diffusion-driven MMIF method is proposed to leverage the information distribution relationship among multi-modality images in the latent space. To better preserve the complementary information from different modalities, a local and global network (LAGN) is suggested. Additionally, a loss strategy is designed to establish robust constraints among diffusion-generated images, original images, and fused images. This strategy supervises the training process and prevents information loss in fused images. The experimental results demonstrate that the proposed method surpasses state-of-the-art image fusion methods in terms of unsupervised metrics on three datasets: MRI/CT, MRI/PET, and MRI/SPECT images. The proposed method successfully captures rich details and color information. Furthermore, 16 doctors and medical students were invited to evaluate the effectiveness of our method in assisting clinical diagnosis and treatment.
PMID:39932643 | DOI:10.1007/s11517-025-03300-6
Double-mix pseudo-label framework: enhancing semi-supervised segmentation on category-imbalanced CT volumes
Int J Comput Assist Radiol Surg. 2025 Feb 11. doi: 10.1007/s11548-024-03281-1. Online ahead of print.
ABSTRACT
PURPOSE: Deep-learning-based supervised CT segmentation relies on fully and densely labeled data, the labeling process of which is time-consuming. In this study, our proposed method aims to improve segmentation performance on CT volumes with limited annotated data by considering category-wise difficulties and distribution.
METHODS: We propose a novel confidence-difficulty weight (CDifW) allocation method that considers confidence levels, balancing the training across different categories, influencing the loss function and volume-mixing process for pseudo-label generation. Additionally, we introduce a novel Double-Mix Pseudo-label Framework (DMPF), which strategically selects categories for image blending based on the distribution of voxel-counts per category and the weight of segmentation difficulty. DMPF is designed to enhance the segmentation performance of categories that are challenging to segment.
RESULT: Our approach was tested on two commonly used datasets: a Congenital Heart Disease (CHD) dataset and a Beyond-the-Cranial-Vault (BTCV) Abdomen dataset. Compared to the SOTA methods, our approach achieved an improvement of 5.1% and 7.0% in Dice score for the segmentation of difficult-to-segment categories on 5% of the labeled data in CHD and 40% of the labeled data in BTCV, respectively.
CONCLUSION: Our method improves segmentation performance in difficult categories within CT volumes by category-wise weights and weight-based mixture augmentation. Our method was validated across multiple datasets and is significant for advancing semi-supervised segmentation tasks in health care. The code is available at https://github.com/MoriLabNU/Double-Mix .
PMID:39932621 | DOI:10.1007/s11548-024-03281-1
Eliminating the second CT scan of dual-tracer total-body PET/CT via deep learning-based image synthesis and registration
Eur J Nucl Med Mol Imaging. 2025 Feb 11. doi: 10.1007/s00259-025-07113-5. Online ahead of print.
ABSTRACT
PURPOSE: This study aims to develop and validate a deep learning framework designed to eliminate the second CT scan of dual-tracer total-body PET/CT imaging.
METHODS: We retrospectively included three cohorts of 247 patients who underwent dual-tracer total-body PET/CT imaging on two separate days (time interval:1-11 days). Out of these, 167 underwent [68Ga]Ga-DOTATATE/[18F]FDG, 50 underwent [68Ga]Ga-PSMA-11/[18F]FDG, and 30 underwent [68Ga]Ga-FAPI-04/[18F]FDG. A deep learning framework was developed that integrates a registration generative adversarial network (RegGAN) with non-rigid registration techniques. This approach allows for the transformation of attenuation-correction CT (ACCT) images from the first scan into pseudo-ACCT images for the second scan, which are then used for attenuation and scatter correction (ASC) of the second tracer PET images. Additionally, the derived registration transform facilitates dual-tracer image fusion and analysis. The deep learning-based ASC PET images were evaluated using quantitative metrics, including mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) across the whole body and specific regions. Furthermore, the quantitative accuracy of PET images was assessed by calculating standardized uptake value (SUV) bias in normal organs and lesions.
RESULTS: The MAE for whole-body pseudo-ACCT images ranged from 97.64 to 112.59 HU across four tracers. The deep learning-based ASC PET images demonstrated high similarity to the ground-truth PET images. The MAE of SUV for whole-body PET images was 0.06 for [68Ga]Ga-DOTATATE, 0.08 for [68Ga]Ga-PSMA-11, 0.06 for [68Ga]Ga-FAPI-04, and 0.05 for [18F]FDG, respectively. Additionally, the median absolute percent deviation of SUV was less than 2.6% for all normal organs, while the mean absolute percent deviation of SUV was less than 3.6% for lesions across four tracers.
CONCLUSION: The proposed deep learning framework, combining RegGAN and non-rigid registration, shows promise in reducing CT radiation dose for dual-tracer total-body PET/CT imaging, with successful validation across multiple tracers.
PMID:39932542 | DOI:10.1007/s00259-025-07113-5
DeepInterAware: Deep Interaction Interface-Aware Network for Improving Antigen-Antibody Interaction Prediction from Sequence Data
Adv Sci (Weinh). 2025 Feb 11:e2412533. doi: 10.1002/advs.202412533. Online ahead of print.
ABSTRACT
Identifying interactions between candidate antibodies and target antigens is a key step in developing effective human therapeutics. The antigen-antibody interaction (AAI) occurs at the structural level, but the limited structure data poses a significant challenge. However, recent studies revealed that structural information can be learned from the vast amount of sequence data, indicating that the interaction prediction can benefit from the abundance of antigen and antibody sequences. In this study, DeepInterAware (deep interaction interface-aware network) is proposed, a framework dynamically incorporating interaction interface information directly learned from sequence data, along with the inherent specificity information of the sequences. Experimental results in interaction prediction demonstrate that DeepInterAware outperforms existing methods and exhibits promising inductive capabilities for predicting interactions involving unseen antigens or antibodies, and transfer capabilities for similar tasks. More notably, DeepInterAware has unique advantages that existing methods lack. First, DeepInterAware can dive into the underlying mechanisms of AAIs, offering the ability to identify potential binding sites. Second, it is proficient in detecting mutations within antigens or antibodies, and can be extended for precise predictions of the binding free energy changes upon mutations. The HER2-targeting antibody screening experiment further underscores DeepInterAware's exceptional capability in identifying binding antibodies for target antigens, establishing it as an important tool for antibody screening.
PMID:39932383 | DOI:10.1002/advs.202412533
ChatExosome: An Artificial Intelligence (AI) Agent Based on Deep Learning of Exosomes Spectroscopy for Hepatocellular Carcinoma (HCC) Diagnosis
Anal Chem. 2025 Feb 11. doi: 10.1021/acs.analchem.4c06677. Online ahead of print.
ABSTRACT
Large language models (LLMs) hold significant promise in the field of medical diagnosis. There are still many challenges in the direct diagnosis of hepatocellular carcinoma (HCC). α-Fetoprotein (AFP) is a commonly used tumor marker for liver cancer. However, relying on AFP can result in missed diagnoses of HCC. We developed an artificial intelligence (AI) agent centered on LLMs, named ChatExosome, which created an interactive and convenient system for clinical spectroscopic analysis and diagnosis. ChatExosome consists of two main components: the first is the deep learning of the Raman fingerprinting of exosomes derived from HCC. Based on a patch-based 1D self-attention mechanism and downsampling, the feature fusion transformer (FFT) was designed to process the Raman spectra of exosomes. It achieved accuracies of 95.8% for cell-derived exosomes and 94.1% for 165 clinical samples, respectively. The second component is the interactive chat agent based on LLM. The retrieval-augmented generation (RAG) method was utilized to enhance the knowledge related to exosomes. Overall, LLM serves as the core of this interactive system, which is capable of identifying users' intentions and invoking the appropriate plugins to process the Raman data of exosomes. This is the first AI agent focusing on exosome spectroscopy and diagnosis, enhancing the interpretability of classification results, enabling physicians to leverage cutting-edge medical research and artificial intelligence techniques to optimize medical decision-making processes, and it shows great potential in intelligent diagnosis.
PMID:39932366 | DOI:10.1021/acs.analchem.4c06677
Correction to "DL 101: Basic Introduction to Deep Learning With Its Application in Biomedical Related Fields"
Stat Med. 2025 Feb 28;44(5):e10349. doi: 10.1002/sim.10349.
NO ABSTRACT
PMID:39932330 | DOI:10.1002/sim.10349
Deep Learning Radiomics Based on MRI for Differentiating Benign and Malignant Parapharyngeal Space Tumors
Laryngoscope. 2025 Feb 11. doi: 10.1002/lary.32043. Online ahead of print.
ABSTRACT
OBJECTIVE: The study aims to establish a pre-academic diagnostic tool based on deep learning and conventional radiomics features to guide the clinical decision-making of parapharyngeal space (PPS) tumors.
METHODS: This retrospective study included 217 patients with PPS tumors, from two medical centers in China from March 1, 2011, to October 1, 2023. The study cohort was divided into a training set (n = 145) and a test set (n = 72). A deep learning (DL) model and conventional radiomics (Rad) model based on neck MRI were constructed to distinguish malignant tumors (MTs) and benign tumors (BTs) of PPS tumors. The deep learning radiomics (DLR) model which integrates deep learning and radiomics features was further developed. The area under the receiver operating characteristic curve (AUC), specificity, and sensitivity were used to evaluate model performance. Decision curve analysis (DCA) was applied to assess the clinical utility.
RESULTS: Compared with the Rad and DL models, the DLR model showed excellent performance in this study, with the highest AUC of 0.899 and 0.821 in the training set and test set, respectively. The DCA curve confirmed the clinical utility of the DLR model in distinguishing the pathological types of PPS tumors.
CONCLUSION: The DLR model demonstrated a high predictive ability in diagnosing MTs and BTs of PPS and could serve as a powerful tool to aid clinical decision-making in the preoperative diagnosis of PPS tumors.
LEVEL OF EVIDENCE: III Laryngoscope, 2025.
PMID:39932109 | DOI:10.1002/lary.32043