Front Physiol. 2024 Dec 13;15:1445258. doi: 10.3389/fphys.2024.1445258. eCollection 2024.

ABSTRACT

INTRODUCTION: This study aimed to develop a deep learning-based method for interpreting magnetic resonance imaging (MRI) scans of temporomandibular joint (TMJ) anterior disc displacement (ADD) and to formulate an automated diagnostic system for clinical practice.

METHODS: The deep learning models were utilized to identify regions of interest (ROI), segment TMJ structures including the articular disc, condyle, glenoid fossa, and articular tubercle, and classify TMJ ADD. The models employed Grad-CAM heatmaps and segmentation annotation diagrams for visual diagnostic predictions and were deployed for clinical application. We constructed four deep-learning models based on the ResNet101_vd framework utilizing an MRI dataset of 618 TMJ cases collected from two hospitals (Hospitals SS and SG) and a dataset of 840 TMJ MRI scans from October 2022 to July 2023. The training and validation datasets included 700 images from Hospital SS, which were used to develop the models. Model performance was assessed using 140 images from Hospital SS (internal validity test) and 140 images from Hospital SG (external validity test). The first model identified the ROI, the second automated the segmentation of anatomical components, and the third and fourth models performed classification tasks based on segmentation and non-segmentation approaches. MRI images were classified into four categories: normal (closed mouth), ADD (closed mouth), normal (open mouth), and ADD (open mouth). Combined findings from open and closed-mouth positions provided conclusive diagnoses. Data augmentation techniques were used to prevent overfitting and enhance model robustness. The models were assessed using performance metrics such as precision, recall, mean average precision (mAP), F1-score, Matthews Correlation Coefficient (MCC), and confusion matrix analysis.

RESULTS: Despite lower performance with Hospital SG's data than Hospital SS's, both achieved satisfactory results. Classification models demonstrated high precision rates above 92%, with the segmentation-based model outperforming the non-segmentation model in overall and category-specific metrics.

DISCUSSION: In summary, our deep learning models exhibited high accuracy in detecting TMJ ADD and provided interpretable, visualized predictive results. These models can be integrated with clinical examinations to enhance diagnostic precision.

PMID:39735724 | PMC:PMC11671476 | DOI:10.3389/fphys.2024.1445258

Oncol Res. 2024 Dec 20;33(1):67-81. doi: 10.32604/or.2024.055921. eCollection 2025.

ABSTRACT

BACKGROUND: Triple-negative breast cancer (TNBC), characterized by its lack of traditional hormone receptors and HER2, presents a significant challenge in oncology due to its poor response to conventional therapies. Autophagy is an important process for maintaining cellular homeostasis, and there are currently autophagy biomarkers that play an effective role in the clinical treatment of tumors. In contrast to targeting protein activity, intervention with protein-protein interaction (PPI) can avoid unrelated crosstalk and regulate the autophagy process with minimal interference pathways.

METHODS: Here, we employed Naive Bayes, Decision Tree, and k-Nearest Neighbors to elucidate the complex PPI network associated with autophagy in TNBC, aiming to uncover novel therapeutic targets. Meanwhile, the candidate proteins interacting with Beclin 2 were initially screened in MDA-MB-231 cells using Beclin 2 as bait protein by immunoprecipitation-mass spectrometry assay, and the interaction relationship was verified by molecular docking and CO-IP experiments after intersection. Colony formation, cellular immunofluorescence, cell scratch and 3-(4,5-Dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT) tests were used to predict the clinical therapeutic effects of manipulating candidate PPI.

RESULTS: By developing three PPI classification models and analyzing over 13,000 datasets, we identified 3733 previously unknown autophagy-related PPIs. Our network analysis revealed the central role of Beclin 2 in autophagy regulation, uncovering its interactions with 39 newly identified proteins. Notably, the CO-IP studies identified the substantial interaction between Beclin 2 and Ubiquilin 1, which was anticipated by our model and discovered in immunoprecipitation-mass spectrometry assay results. Subsequently, in vitro investigations showed that overexpressing Beclin 2 increased Ubiquilin 1, promoted autophagy-dependent cell death, and inhibited proliferation and metastasis in MDA-MB-231 cells.

CONCLUSIONS: This study not only enhances our understanding of autophagy regulation in TNBC but also identifies the Beclin 2-Ubiquilin 1 axis as a promising target for precision therapy. These findings open new avenues for drug discovery and offer inspiration for more effective treatments for this aggressive cancer subtype.

PMID:39735677 | PMC:PMC11671618 | DOI:10.32604/or.2024.055921

Bioinform Biol Insights. 2024 Dec 24;18:11779322241249562. doi: 10.1177/11779322241249562. eCollection 2024.

ABSTRACT

In recent years, several machine learning (ML) approaches have been proposed to predict gene expression signal and chromatin features from the DNA sequence alone. These models are often used to deduce and, to some extent, assess putative new biological insights about gene regulation, and they have led to very interesting advances in regulatory genomics. This article reviews a selection of these methods, ranging from linear models to random forests, kernel methods, and more advanced deep learning models. Specifically, we detail the different techniques and strategies that can be used to extract new gene-regulation hypotheses from these models. Furthermore, because these putative insights need to be validated with wet-lab experiments, we emphasize that it is important to have a measure of confidence associated with the extracted hypotheses. We review the procedures that have been proposed to measure this confidence for the different types of ML models, and we discuss the fact that they do not provide the same kind of information.

PMID:39735654 | PMC:PMC11672376 | DOI:10.1177/11779322241249562

Front Oncol. 2024 Dec 13;14:1458374. doi: 10.3389/fonc.2024.1458374. eCollection 2024.

ABSTRACT

BACKGROUND: The aim of this study is to develop deep learning models based on 18F-fluorodeoxyglucose positron emission tomography/computed tomographic (18F-FDG PET/CT) images for predicting individual epidermal growth factor receptor (EGFR) mutation status in lung adenocarcinoma (LUAD).

METHODS: We enrolled 430 patients with non-small-cell lung cancer from two institutions in this study. The advanced Inception V3 model to predict EGFR mutations based on PET/CT images and developed CT, PET, and PET + CT models was used. Additionally, each patient's clinical characteristics (age, sex, and smoking history) and 18 CT features were recorded and analyzed. Univariate and multivariate regression analyses identified the independent risk factors for EGFR mutations, and a clinical model was established. The performance using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, recall, and F1-value was evaluated. The DeLong test was used to compare the predictive performance across various models.

RESULTS: Among these four models, deep learning models based on CT and PET + CT exhibit the same predictive performance, followed by PET and the clinical model. The AUC values for CT, PET, PET + CT, and clinical models in the training set are 0.933 (95% CI, 0.922-0.943), 0.895 (95% CI, 0.882-0.907), 0.931 (95% CI, 0.921-0.942), and 0.740 (95% CI, 0.685-0.796), respectively; whereas those in the testing set are:0.921 (95% CI, 0.904-0.938), 0.876 (95% CI, 0.855-0.897), 0.921 (95% CI, 0.904-0.937), and 0.721 (95% CI, 0.629-0.814), respectively. The DeLong test results confirm that all deep learning models are superior to clinical one.

CONCLUSION: The PET/CT images based on trained CNNs effectively predict EGFR and non-EGFR mutations in LUAD. The deep learning predictive models could guide treatment options.

PMID:39735601 | PMC:PMC11671303 | DOI:10.3389/fonc.2024.1458374

Front Oncol. 2024 Dec 13;14:1414609. doi: 10.3389/fonc.2024.1414609. eCollection 2024.

ABSTRACT

OBJECTIVES: The accurate assessment of lymph node metastasis (LNM) can facilitate clinical decision-making on radiotherapy or radical hysterectomy (RH) in cervical adenocarcinoma (AC)/adenosquamous carcinoma (ASC). This study aims to develop a deep learning radiomics nomogram (DLRN) to preoperatively evaluate LNM in cervical AC/ASC.

MATERIALS AND METHODS: A total of 652 patients from a multicenter were enrolled and randomly allocated into primary, internal, and external validation cohorts. The radiomics features were extracted from axial T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and contrast-enhanced T1-weighted imaging (CE-T1WI). The DL features from T2WI, DWI, and CE-T1WI were exported from Resnet 34, which was pretrained by 14 million natural images of the ImageNet dataset. The radscore (RS) and DL score (DLS) were independently obtained after repeatability test, Pearson correlation coefficient (PCC), minimum redundancy maximum relevance (MRMR), and least absolute shrinkage and selection operator (LASSO) algorithm performed on the radiomics and DL feature sets. The DLRN was then developed by integrating the RS, DLS, and independent clinicopathological factors for evaluating the LNM in cervical AC/ASC.

RESULTS: The nomogram of DLRN-integrated FIGO stage, menopause, RS, and DLS achieved AUCs of 0.79 (95% CI, 0.74-0.83), 0.87 (95% CI, 0.81-0.92), and 0.86 (95% CI, 0.79-0.91) in the primary, internal, and external validation cohorts. Compared with the RS, DLS, and clinical models, DLRN had a significant higher AUC for evaluating LNM (all P < 0.005).

CONCLUSIONS: The nomogram of DLRN can accurately evaluate LNM in cervical AC/ASC.

PMID:39735600 | PMC:PMC11671353 | DOI:10.3389/fonc.2024.1414609

Front Immunol. 2024 Dec 13;15:1482020. doi: 10.3389/fimmu.2024.1482020. eCollection 2024.

ABSTRACT

OBJECTIVE: To explore the value of combined radiomics and deep learning models using different machine learning algorithms based on mammography (MG) and magnetic resonance imaging (MRI) for predicting axillary lymph node metastasis (ALNM) in breast cancer (BC). The objective is to provide guidance for developing scientifically individualized treatment plans, assessing prognosis, and planning preoperative interventions.

METHODS: A retrospective analysis was conducted on clinical and imaging data from 270 patients with BC confirmed by surgical pathology at the Third Hospital of Shanxi Medical University between November 2022 and April 2024. Multiple sequence images from MG and MRI were selected, and regions of interest in the lesions were delineated. Radiomics and deep learning (3D-Resnet18) features were extracted and fused. The samples were randomly divided into training and test sets in a 7:3 ratio. Dimensionality reduction and feature selection were performed using the least absolute shrinkage and selection operator (LASSO) regression model, and other methods. Various machine learning algorithms were used to construct radiomics, deep learning, and combined models. These models were visualized and evaluated for performance using receiver operating characteristic curves, area under the curve (AUC), calibration curves, and decision curves.

RESULTS: The highest AUCs in the test set were achieved using radiomics-logistic regression (AUC = 0.759), deep learning-multilayer perceptron (MLP) (AUC = 0.712), and combined-MLP models (AUC = 0.846). The MLP model demonstrated strong classification performance, with the combined model (AUC = 0.846) outperforming both the radiomics (AUC = 0.756) and deep learning (AUC = 0.712) models.

CONCLUSION: The multimodal radiomics and deep learning models developed in this study, incorporating various machine learning algorithms, offer significant value for the preoperative prediction of ALNM in BC.

PMID:39735531 | PMC:PMC11671510 | DOI:10.3389/fimmu.2024.1482020

Health Care Sci. 2024 Nov 28;3(6):438-455. doi: 10.1002/hcs2.121. eCollection 2024 Dec.

ABSTRACT

BACKGROUND: Skin cancer poses a significant global health threat, with early detection being essential for successful treatment. While deep learning algorithms have greatly enhanced the categorization of skin lesions, the black-box nature of many models limits interpretability, posing challenges for dermatologists.

METHODS: To address these limitations, SkinSage XAI utilizes advanced explainable artificial intelligence (XAI) techniques for skin lesion categorization. A data set of around 50,000 images from the Customized HAM10000, selected for diversity, serves as the foundation. The Inception v3 model is used for classification, supported by gradient-weighted class activation mapping and local interpretable model-agnostic explanations algorithms, which provide clear visual explanations for model outputs.

RESULTS: SkinSage XAI demonstrated high performance, accurately categorizing seven types of skin lesions-dermatofibroma, benign keratosis, melanocytic nevus, vascular lesion, actinic keratosis, basal cell carcinoma, and melanoma. It achieved an accuracy of 96%, with precision at 96.42%, recall at 96.28%, f 1 score at 96.14%, and an area under the curve of 99.83%.

CONCLUSIONS: SkinSage XAI represents a significant advancement in dermatology and artificial intelligence by bridging gaps in accuracy and explainability. The system provides transparent, accurate diagnoses, improving decision-making for dermatologists and potentially enhancing patient outcomes.

PMID:39735286 | PMC:PMC11671215 | DOI:10.1002/hcs2.121

Health Care Sci. 2024 Dec 15;3(6):456-474. doi: 10.1002/hcs2.119. eCollection 2024 Dec.

ABSTRACT

BACKGROUND: Pneumothorax is a medical emergency caused by the abnormal accumulation of air in the pleural space-the potential space between the lungs and chest wall. On 2D chest radiographs, pneumothorax occurs within the thoracic cavity and outside of the mediastinum, and we refer to this area as "lung + space." While deep learning (DL) has increasingly been utilized to segment pneumothorax lesions in chest radiographs, many existing DL models employ an end-to-end approach. These models directly map chest radiographs to clinician-annotated lesion areas, often neglecting the vital domain knowledge that pneumothorax is inherently location-sensitive.

METHODS: We propose a novel approach that incorporates the lung + space as a constraint during DL model training for pneumothorax segmentation on 2D chest radiographs. To circumvent the need for additional annotations and to prevent potential label leakage on the target task, our method utilizes external datasets and an auxiliary task of lung segmentation. This approach generates a specific constraint of lung + space for each chest radiograph. Furthermore, we have incorporated a discriminator to eliminate unreliable constraints caused by the domain shift between the auxiliary and target datasets.

RESULTS: Our results demonstrated considerable improvements, with average performance gains of 4.6%, 3.6%, and 3.3% regarding intersection over union, dice similarity coefficient, and Hausdorff distance. These results were consistent across six baseline models built on three architectures (U-Net, LinkNet, or PSPNet) and two backbones (VGG-11 or MobileOne-S0). We further conducted an ablation study to evaluate the contribution of each component in the proposed method and undertook several robustness studies on hyper-parameter selection to validate the stability of our method.

CONCLUSIONS: The integration of domain knowledge in DL models for medical applications has often been underemphasized. Our research underscores the significance of incorporating medical domain knowledge about the location-specific nature of pneumothorax to enhance DL-based lesion segmentation and further bolster clinicians' trust in DL tools. Beyond pneumothorax, our approach is promising for other thoracic conditions that possess location-relevant characteristics.

PMID:39735285 | PMC:PMC11671217 | DOI:10.1002/hcs2.119

World J Gastroenterol. 2024 Dec 28;30(48):5111-5129. doi: 10.3748/wjg.v30.i48.5111.

ABSTRACT

BACKGROUND: Wireless capsule endoscopy (WCE) has become an important noninvasive and portable tool for diagnosing digestive tract diseases and has been propelled by advancements in medical imaging technology. However, the complexity of the digestive tract structure, and the diversity of lesion types, results in different sites and types of lesions distinctly appearing in the images, posing a challenge for the accurate identification of digestive tract diseases.

AIM: To propose a deep learning-based lesion detection model to automatically identify and accurately label digestive tract lesions, thereby improving the diagnostic efficiency of doctors, and creating significant clinical application value.

METHODS: In this paper, we propose a neural network model, WCE_Detection, for the accurate detection and classification of 23 classes of digestive tract lesion images. First, since multicategory lesion images exhibit various shapes and scales, a multidetection head strategy is adopted in the object detection network to increase the model's robustness for multiscale lesion detection. Moreover, a bidirectional feature pyramid network (BiFPN) is introduced, which effectively fuses shallow semantic features by adding skip connections, significantly reducing the detection error rate. On the basis of the above, we utilize the Swin Transformer with its unique self-attention mechanism and hierarchical structure in conjunction with the BiFPN feature fusion technique to enhance the feature representation of multicategory lesion images.

RESULTS: The model constructed in this study achieved an mAP50 of 91.5% for detecting 23 lesions. More than eleven single-category lesions achieved an mAP50 of over 99.4%, and more than twenty lesions had an mAP50 value of over 80%. These results indicate that the model outperforms other state-of-the-art models in the end-to-end integrated detection of human digestive tract lesion images.

CONCLUSION: The deep learning-based object detection network detects multiple digestive tract lesions in WCE images with high accuracy, improving the diagnostic efficiency of doctors, and demonstrating significant clinical application value.

PMID:39735271 | PMC:PMC11612692 | DOI:10.3748/wjg.v30.i48.5111

Front Artif Intell. 2024 Dec 13;7:1453931. doi: 10.3389/frai.2024.1453931. eCollection 2024.

ABSTRACT

Many industries utilize deep learning methods to increase efficiency and reduce costs. One of these methods, image segmentation, is used for object detection and recognition in localization and mapping. Segmentation models are trained using labeled datasets; however, manually creating datasets for every application, including deep-level mining, is time-consuming and typically expensive. Recently, many papers have shown that using synthetic datasets (digital recreations of real-world scenes) for training produces highly-accurate segmentation models. This paper proposes a synthetic segmentation dataset generator using a 3D modeling framework and raycaster. The generator was applied to a deep-level mining case study and produced a dataset containing labeled images of scenes typically found in this environment, therefore removing the requirement to create the dataset manually. Validation showed high accuracy segmentation after model training using the generated dataset (compared to other applications that use real-world datasets). Furthermore, the generator can be customized to produce datasets for many other applications.

PMID:39735233 | PMC:PMC11672338 | DOI:10.3389/frai.2024.1453931

J Cheminform. 2024 Dec 29;16(1):147. doi: 10.1186/s13321-024-00938-6.

ABSTRACT

Predicting protein-ligand binding affinity is essential for understanding protein-ligand interactions and advancing drug discovery. Recent research has demonstrated the advantages of sequence-based models and graph-based models. In this study, we present a novel hybrid multimodal approach, DeepTGIN, which integrates transformers and graph isomorphism networks to predict protein-ligand binding affinity. DeepTGIN is designed to learn sequence and graph features efficiently. The DeepTGIN model comprises three modules: the data representation module, the encoder module, and the prediction module. The transformer encoder learns sequential features from proteins and protein pockets separately, while the graph isomorphism network extracts graph features from the ligands. To evaluate the performance of DeepTGIN, we compared it with state-of-the-art models using the PDBbind 2016 core set and PDBbind 2013 core set. DeepTGIN outperforms these models in terms of R, RMSE, MAE, SD, and CI metrics. Ablation studies further demonstrate the effectiveness of the ligand features and the encoder module. The code is available at: https://github.com/zhc-moushang/DeepTGIN . SCIENTIFIC CONTRIBUTION: DeepTGIN is a novel hybrid multimodal deep learning model for predict protein-ligand binding affinity. The model combines the Transformer encoder to extract sequence features from protein and protein pocket, while integrating graph isomorphism networks to capture features from the ligand. This model addresses the limitations of existing methods in exploring protein pocket and ligand features.

PMID:39734235 | DOI:10.1186/s13321-024-00938-6

Radiother Oncol. 2024 Dec 27:110699. doi: 10.1016/j.radonc.2024.110699. Online ahead of print.

ABSTRACT

BACKGROUND AND PURPOSE: Accurate segmentation of the clinical target volume (CTV) is essential to deliver an effective radiation dose to tumor tissues in cervical cancer radiotherapy. Also, although automated CTV segmentation can reduce oncologists' workload, challenges persist due to the microscopic spread of tumor cells undetectable in CT imaging, low-intensity contrast between organs, and inter-observer variability. This study aims to develop and validate a multi-task feature fusion network (MTF-Net) that uses distance-based information to enhance CTV segmentation accuracy.

MATERIALS AND METHODS: We developed a dual-branch, end-to-end MTF-Net designed to address the challenges in cervical cancer CTV segmentation. The MTF-Net architecture consists of a shared encoder and two parallel decoders, one generating a distance location information map (Dimg) and the other producing CTV segmentation masks. To enhance segmentation accuracy, we introduced a distance information attention fusion module that integrates features from the Dimg into the CTV segmentation process, thus optimizing target delineation. The datasets for this study were from three centers. Data from two centers were used for model training and internal validation, and that of the third center was used as an independent dataset for external testing. To benchmark performance, we also compared MTF-Net to commercial segmentation software in a clinical setting.

RESULTS: MTF-Net achieved an average dice score of 84.67% on internal and 77.51% on external testing datasets. Compared with commercial software, MTF-Net demonstrated superior performance across several metrics, including Dice score, positive predictive value, and 95% Hausdorff distance, with the exception of sensitivity.

CONCLUSIONS: This study indicates that MTF-Net outperforms existing state-of-the-art segmentation methods and commercial software, demonstrating its potential effectiveness for clinical applications in cervical cancer radiotherapy planning.

PMID:39733971 | DOI:10.1016/j.radonc.2024.110699

J Gynecol Obstet Hum Reprod. 2024 Dec 27:102903. doi: 10.1016/j.jogoh.2024.102903. Online ahead of print.

ABSTRACT

In-vitro fertilization (IVF) has been a transformative advancement in assisted reproductive technology. However, success rates remain suboptimal, with only about one-third of cycles resulting in pregnancy and fewer leading to live births. This narrative review explores the potential of artificial intelligence (AI), machine learning (ML), and deep learning (DL) to enhance various stages of the IVF process. Personalization of ovarian stimulation protocols, gamete selection, and embryo annotation and selection are critical areas where AI may benefit significantly. AI-driven tools can analyze vast datasets to predict optimal stimulation protocols, potentially improving oocyte quality and fertilization rates. In sperm and oocyte quality assessment, AI can offer precise, objective analyses, reducing subjectivity and standardizing evaluations. In embryo selection, AI can analyze time-lapse imaging and morphological data to support the prediction of embryo viability, potentially aiding implantation outcomes. However, the role of AI in improving clinical outcomes remains to be confirmed by large-scale, well-designed clinical trials. Additionally, AI has the potential to enhance quality control and workflow optimization within IVF laboratories by continuously monitoring key performance indicators (KPIs) and facilitating efficient resource utilization. Ethical considerations, including data privacy, algorithmic bias, and fairness, are paramount for the responsible implementation of AI in IVF. Future research should prioritize validating AI tools in diverse clinical settings, ensuring their applicability and reliability. Collaboration among AI experts, clinicians, and embryologists is essential to drive innovation and improve outcomes in assisted reproduction. AI's integration into IVF holds promise for advancing patient care, but its clinical potential requires careful evaluation and ongoing refinement.

PMID:39733809 | DOI:10.1016/j.jogoh.2024.102903

Kidney Int. 2024 Dec 27:S0085-2538(24)00923-2. doi: 10.1016/j.kint.2024.12.007. Online ahead of print.

ABSTRACT

The response of the kidney after induction treatment is one of the determinants of prognosis in lupus nephritis, but effective predictive tools are lacking. Here, we sought to apply deep learning approaches on kidney biopsies for treatment response prediction in lupus nephritis. Patients who received cyclophosphamide or mycophenolate mofetil as induction treatment were included and the primary outcome was 12-month treatment response, complete response defined as 24h urinary protein under 0.5 g with normal estimated glomerular filtration rate or within 10% of normal range. The model development cohort included 245 patients (880 digital slides), and the external test cohort had 71 patients (258 digital slides). Deep learning models were trained independently on hematoxylin and eosin, periodic acid-Schiff, periodic Schiff-methenamine silver and Masson's trichrome stained slides at multiple magnifications and integrated to predict the primary outcome of complete response to therapy at 12 months. Single-stain models showed area under the curves of 0.813, 0.841, 0.823, and 0.862, respectively. Further, integration of the four models into a multi-stain model achieved area under the curves of 0.901 and 0.840 on internal validation and external testing, respectively, which outperformed conventional clinicopathologic parameters including estimated glomerular filtration rate, chronicity index and reduction in proteinuria at three months. Decisive features uncovered by visualization uncovered for model prediction included tertiary lymphoid structures, glomerulosclerosis, interstitial fibrosis and tubular atrophy. Our study demonstrated the feasibility of utilizing deep learning on kidney pathology to predict treatment response for lupus patients. Further validation is required before the model could be implemented for risk stratification and to aid in making therapeutic decisions in clinical practice.

PMID:39733792 | DOI:10.1016/j.kint.2024.12.007

Comput Methods Programs Biomed. 2024 Dec 25;260:108575. doi: 10.1016/j.cmpb.2024.108575. Online ahead of print.

ABSTRACT

BACKGROUND AND OBJECTIVE: Sparse-view computed tomography (CT) speeds up scanning and reduces radiation exposure in medical diagnosis. However, when the projection views are severely under-sampled, deep learning-based reconstruction methods often suffer from over-smoothing of the reconstructed images due to the lack of high-frequency information. To address this issue, we introduce frequency domain information into the popular projection-image domain reconstruction, proposing a Tri-Domain sparse-view CT reconstruction model based on Sparse Transformer (TD-STrans).

METHODS: TD-STrans integrates three essential modules: the projection recovery module completes the sparse-view projection, the Fourier domain filling module mitigates artifacts and over-smoothing by filling in missing high-frequency details; the image refinement module further enhances and preserves image details. Additionally, a multi-domain joint loss function is designed to simultaneously enhance the reconstruction quality in the projection domain, image domain, and frequency domain, thereby further improving the preservation of image details.

RESULTS: The results of simulation experiments on the lymph node dataset and real experiments on the walnut dataset consistently demonstrate the effectiveness of TD-STrans in artifact removal, suppression of over-smoothing, and preservation of structural fidelity.

CONCLUSION: The reconstruction results of TD-STrans indicate that sparse transformer across multiple domains can alleviate over-smoothing and detail loss caused by reduced views, offering a novel solution for ultra-sparse-view CT imaging.

PMID:39733746 | DOI:10.1016/j.cmpb.2024.108575

Spectrochim Acta A Mol Biomol Spectrosc. 2024 Dec 22;330:125653. doi: 10.1016/j.saa.2024.125653. Online ahead of print.

ABSTRACT

Wheat flour quality, determined by factors such as protein and moisture content, is crucial in food production. Traditional methods for analyzing these parameters, though precise, are time-consuming and impractical for large-scale operations. This study presents a lightweight convolutional neural network designed for real-time wheat flour quality monitoring using near-infrared spectroscopy. The model incorporates Ghost bottlenecks, external attention modules, and the Kolmogorov-Arnold network to enhance feature extraction and improve prediction accuracy. Testing results demonstrate high predictive performance with R2 values of 0.9653 (RMSE: 0.2886 g/100 g, RPD: 5.8981) for protein and 0.9683 (RMSE: 0.3061 g/100 g, RPD: 5.1046) for moisture content. The model's robustness across diverse samples and its suitability for online applications make it a promising tool for efficient and non-destructive quality control in the food industry.

PMID:39733712 | DOI:10.1016/j.saa.2024.125653

Spectrochim Acta A Mol Biomol Spectrosc. 2024 Dec 27;330:125679. doi: 10.1016/j.saa.2024.125679. Online ahead of print.

ABSTRACT

Raman spectroscopy is a powerful and non-invasive analytical method for determining the chemical composition and molecular structure of a wide range of materials, including complex biological tissues. However, the captured signals typically suffer from interferences manifested as noise and baseline, which need to be removed for successful data analysis. Effective baseline correction is critical in quantitative analysis, as it may impact peak signature derivation. Current baseline correction methods can be labor-intensive and may require extensive parameter adjustment depending on the input spectrum characteristics. In contrast, deep learning-based baseline correction models trained across various materials, offer a promising and more versatile alternative. This study reports an approach to manually identify the ground-truth baselines for eight different biological materials through extensively tuning the parameters of three classical baseline correction methods, Modified Multi-Polynomial Fit (Modpoly), Improved Modified Multi-Polynomial Fitting (IModpoly), and Adaptive Iteratively Reweighted Penalized Least Squares (airPLS), and combining the outputs to best fit the training data. We designed a one-dimensional Transformer (1dTrans) tailored to fit Raman spectral data for estimating their baselines, and evaluated its performance against convolutional neural network (CNN), ResUNet, and three aforementioned parametric methods. The 1dTrans model achieved lower mean absolute error (MAE) and spectral angle mapper (SAM) scores when compared to the other methods in both development and evaluation of the manually labeled original raw Raman spectra, highlighting the effectiveness of the method in Raman spectra pre-processing.

PMID:39733708 | DOI:10.1016/j.saa.2024.125679

Semergen. 2024 Dec 28;51(5):102434. doi: 10.1016/j.semerg.2024.102434. Online ahead of print.

ABSTRACT

INTRODUCTION: Orphan diseases (OD) are rare but collectively common, presenting challenges such as late diagnoses, disease progression, and limited therapeutic options. Recently, artificial intelligence (AI) has gained interest in the research of these diseases.

OBJECTIVE: To synthesize the available evidence on the use of AI in the comprehensive approach to orphan diseases.

METHODS: An exploratory systematic review of the Scoping Review type was conducted in PubMed, Bireme, and Scopus from 2019 to 2024.

RESULTS: fifty-six articles were identified, with 21.4% being experimental studies; 28 documents did not specify an OD, 8 documents focused primarily on genetic diseases; 53.57% focused on diagnosis, and 36 different algorithms were identified.

CONCLUSIONS: The information found shows the development of AI algorithms in different clinical settings, confirming the potential benefits in diagnosis times, therapeutic options, and greater awareness among health professionals.

PMID:39733637 | DOI:10.1016/j.semerg.2024.102434

Comput Biol Med. 2024 Dec 28;186:109573. doi: 10.1016/j.compbiomed.2024.109573. Online ahead of print.

ABSTRACT

Classification of viruses carries important implications in terms of understanding their evolution and the designing of interventions. This study introduces ViroNia as a novel LSTM-based system specifically meant for high-accuracy classification of viral proteins. Although originally developed for generative tasks, LSTM architectures have been found to be highly efficient for classification tasks as well; the model demonstrates this capability. It outperforms the deep architectures, such as Simple RNN, GRU, 1d CNN and Bidirectional LSTM, with the advantage of using pairwise sequence similarity and efficient data handling. ViroNia, with a dataset of 2250 protein sequences from both the NCBI and BVBRC databases, shows great performance at accuracy levels of 99.7 % and 99.6 % for broad as well as detail-level classifications, respectively. Cross-validation was carried out on the data provided for the fivefold strategy, achieving average accuracies of 92.29 % (±1.55 %) and 90.31 % (±5.41 %), respectively, at both the broad and detail level. The architecture allows for real-time data processing and automatic feature extraction, addressing the scalability limitations faced by BLAST (Basic Local Alignment Search Tool). The comparative analysis revealed that, although existing deep learning models share similar training parameters, ViroNia significantly enhanced classification outcomes. It finds specific applications in those areas that demand real-time analysis and learning on extra viral protein datasets, and hence, contributes broadly to ongoing viral research.

PMID:39733555 | DOI:10.1016/j.compbiomed.2024.109573

Comput Biol Med. 2024 Dec 28;186:109608. doi: 10.1016/j.compbiomed.2024.109608. Online ahead of print.

ABSTRACT

BACKGROUND: Electroencephalogram (EEG) signals-based motor kinematics prediction (MKP) has been an active area of research to develop Brain-computer interface (BCI) systems such as exosuits, prostheses, and rehabilitation devices. However, EEG source imaging (ESI) based kinematics prediction is sparsely explored in the literature.

METHOD: In this study, pre-movement EEG features are utilized to predict three-dimensional (3D) hand kinematics for the grasp-and-lift motor task. A public dataset, WAY-EEG-GAL, is utilized for MKP analysis. In particular, sensor-domain (EEG data) and source-domain (ESI data) based features from the frontoparietal region are explored for MKP. Deep learning-based models are explored to achieve efficient kinematics decoding. Various time-lagged and window sizes are analyzed for hand kinematics prediction. Subsequently, intra-subject and inter-subject MKP analysis is performed to investigate the subject-specific and subject-independent motor-learning capabilities of the neural decoders. The Pearson correlation coefficient (PCC) is used as the performance metric for kinematics trajectory decoding.

RESULTS: The rEEGNet neural decoder achieved the best performance with sensor-domain and source-domain features with the time lag and window size of 100ms and 450ms, respectively. The highest mean PCC values of 0.790, 0.795, and 0.637 are achieved using sensor-domain features, while 0.769, 0.777, and 0.647 are achieved using source-domain features in x, y, and z-directions, respectively.

CONCLUSION: This study explores the feasibility of trajectory prediction using EEG sensor-domain and source-domain features for the grasp-and-lift task. Furthermore, inter-subject trajectory estimation is performed using the proposed deep learning decoder with EEG source domain features.

PMID:39733553 | DOI:10.1016/j.compbiomed.2024.109608