Sci Rep. 2024 Sep 30;14(1):22729. doi: 10.1038/s41598-024-73863-1.

ABSTRACT

Enhancing deep learning performance requires extensive datasets. Centralized training raises concerns about data ownership and security. Additionally, large models are often unsuitable for hospitals due to their limited resource capacities. Federated learning (FL) has been introduced to address these issues. However, FL faces challenges such as vulnerability to attacks, non-IID data, reliance on a central server, high communication overhead, and suboptimal model aggregation. Furthermore, FL is not optimized for realistic hospital database environments, where data are dynamically accumulated. To overcome these limitations, we propose federated influencer learning (FIL) as a secure and efficient collaborative learning paradigm. Unlike the server-client model of FL, FIL features an equal-status structure among participants, with an administrator overseeing the overall process. FIL comprises four stages: local training, qualification, screening, and influencing. Local training is similar to vanilla FL, except for the optional use of a shared dataset. In the qualification stage, participants are classified as influencers or followers. During the screening stage, the integrity of the logits from the influencer is examined. If the integrity is confirmed, the influencer shares their knowledge with the others. FIL is more secure than FL because it eliminates the need for model-parameter transactions, central servers, and generative models. Additionally, FIL supports model-agnostic training. These features make FIL particularly promising for fields such as healthcare, where maintaining confidentiality is crucial. Our experiments demonstrated the effectiveness of FIL, which outperformed several FL methods on large medical (X-ray, MRI, and PET) and natural (CIFAR-10) image dataset in a dynamically accumulating database environment, with consistently higher precision, recall, Dice score, and lower standard deviation between participants. In particular, in the PET dataset, FIL achieved about a 40% improvement in Dice score and recall.

PMID:39349569 | DOI:10.1038/s41598-024-73863-1

Sci Rep. 2024 Sep 30;14(1):22654. doi: 10.1038/s41598-024-73357-0.

ABSTRACT

Current research in deep learning, which is widely used in mineral prospectivity prediction, focuses on obtaining high-performance models to predict mineral resources. However, because the network structure and depth of different algorithms differ, there are some differences in the correlation between the spatial pattern of ore-generating geological big data and the spatial location of discovered ore deposits; this causes instability in the prediction. To solve this problem, this paper proposes the use of ensemble learning to synthesize convolutional neural network algorithms and self-attention mechanism algorithms for mineral prospectivity prediction. In this study, 14 factors related to gold mineralization were selected, 10 types of geochemical exploration data (Au, Ag, As, Cu, Pb, Zn, Hg, Sb, W, and Mo) and 4 geological factors (ductile shear zones, brittle fault zones, mineralization-alteration body zones, and metamorphic quartz sandstone zones). Six classical convolutional neural network models (MobileNet V2, ResNet 50, VGG 16, AlexNet, LeNet, and VIT) were used to extract the features of the metallogenic factors. After training, a network model with an accuracy over 94% was obtained. Then, the mineral prospectivity of an unknown area was predicted. The models were evaluated according to their accuracy. Using these results, ensemble learning was performed, areas with high potential were obtained, and the prospectivity prediction map was drawn. This map provides guidance for gold exploration in the Bawanggou mine area of the northern Hanyin gold orefield, South Qinling, China. This comprehensive method can effectively leverage the advantages of various models, fully extract the internal relationships of deep-level mineralization, and has extremely high extensibility. The calculated results can be made more scientific and stable by adding more mineralization factors and introducing an algorithm with the new structure in the future.

PMID:39349559 | DOI:10.1038/s41598-024-73357-0

Sci Rep. 2024 Sep 30;14(1):22683. doi: 10.1038/s41598-024-71757-w.

ABSTRACT

Antibiotic resistance in bacterial pathogens is a major threat to global health, exacerbated by the misuse of antibiotics. In hospital practice, results of bacterial cultures and antibiograms can take several days. Meanwhile, prescribing an empirical antimicrobial treatment is challenging, as clinicians must balance the antibiotic spectrum against the expected probability of susceptibility. We present here a proof of concept study of a machine learning-based system that predicts the probability of antimicrobial susceptibility and explains the contribution of the different cofactors in hospitalized patients, at four different stages prior to the antibiogram (sampling, direct examination, positive culture, and species identification), using only historical bacterial ecology data that can be easily collected from any laboratory information system (LIS) without GDPR restrictions once the data have been anonymised. A comparative analysis of different state-of-the-art machine learning and probabilistic methods was performed using 44,026 instances over 7 years from the Hôpital Européen Marseille, France. Our results show that multilayer dense neural networks and Bayesian models are suitable for early prediction of antibiotic susceptibility, with AUROCs reaching 0.88 at the positive culture stage and 0.92 at the species identification stage, and even 0.82 and 0.92, respectively, for the least frequent situations. Perspectives and potential clinical applications of the system are discussed.

PMID:39349551 | DOI:10.1038/s41598-024-71757-w

Sci Rep. 2024 Sep 30;14(1):22648. doi: 10.1038/s41598-024-72648-w.

ABSTRACT

The heterogeneity of Parkinson's disease (PD) presents considerable challenges for accurate diagnosis, particularly during early-stage disease, when the symptoms may be extremely subtle. This study aimed to assess the accuracy of a convolutional neural network (CNN) technique based on the 6-min walk test (6MWT) measured using wearable sensors to distinguish patients with early-stage PD (n = 78) from healthy controls (n = 50). The participants wore six sensors, and performed the 6MWT. The time-series data were converted into new images. The results revealed that the gyroscopic vertical component of the lumbar spine displayed the highest classification accuracy of 83.5%, followed by those of the thoracic spine (83.1%) and right thigh (79.5%) segment. These findings suggest that the 6MWT and CNN models may facilitate earlier diagnosis and monitoring of PD symptoms, enabling clinicians to provide timely treatment during the critical transition from normal to pathologic gait patterns.

PMID:39349539 | DOI:10.1038/s41598-024-72648-w

Spectrochim Acta A Mol Biomol Spectrosc. 2024 Sep 24;326:125212. doi: 10.1016/j.saa.2024.125212. Online ahead of print.

ABSTRACT

Accurate, fast and non-destructive identification of varieties of common bean seeds is important for the cultivation and efficient utilization of common beans. This study is based on hyperspectral and deep learning to identify the varieties of common bean seeds non-destructively. In this study, the average spectrum of 3078 hyperspectral images from 500 varieties was obtained after image segmentation and sensitive region extraction, and the Synthetic Minority Over-sampling Technique (SMOTE) was used to achieve the equilibrium of the samples of various varieties. A one-dimensional convolutional neural network model (IResCNN) incorporating Inception module and residual structure was proposed to identify seed varieties, and Support Vector Machine (SVM), K-Nearest Neighbor (KNN), VGG19, AlexNet, ResNet50 were established to compare the identification effect. After analyzing the effects of multiple spectral preprocessing methods on the model, the study selected Savitzky-Golay smoothing correction (SG) for spectral preprocessing and extracted 66 characteristic wavelengths using Successive Projections Algorithm (SPA) as inputs to the discriminative model. Ultimately, the IResCNN model achieved the highest accuracy of 93.06 % on the test set, indicating that hyperspectral technology can accurately identify bean varieties, and the study provides a correct method of thinking for the non-destructive classification of multi-species small-sample bean varieties.

PMID:39348737 | DOI:10.1016/j.saa.2024.125212

Proc Natl Acad Sci U S A. 2024 Oct 8;121(41):e2416358121. doi: 10.1073/pnas.2416358121. Epub 2024 Sep 30.

NO ABSTRACT

PMID:39348547 | DOI:10.1073/pnas.2416358121

PLoS Comput Biol. 2024 Sep 30;20(9):e1012497. doi: 10.1371/journal.pcbi.1012497. Online ahead of print.

ABSTRACT

Accurate prediction of epidemics is pivotal for making well-informed decisions for the control of infectious diseases, but addressing heterogeneity in the system poses a challenge. In this study, we propose a novel modeling framework integrating the spatio-temporal heterogeneity of susceptible individuals into homogeneous models, by introducing a continuous recruitment process for the susceptibles. A neural network approximates the recruitment rate to develop a Universal Differential Equations (UDE) model. Simultaneously, we pre-set a specific form for the recruitment rate and develop a mechanistic model. Data from a COVID Omicron variant outbreak in Shanghai are used to train the UDE model using deep learning methods and to calibrate the mechanistic model using MCMC methods. Subsequently, we project the attack rate and peak of new infections for the first Omicron wave in China after the adjustment of the dynamic zero-COVID policy. Our projections indicate an attack rate and a peak of new infections of 80.06% and 3.17% of the population, respectively, compared with the homogeneous model's projections of 99.97% and 32.78%, thus providing an 18.6% improvement in the prediction accuracy based on the actual data. Our simulations demonstrate that heterogeneity in the susceptibles decreases herd immunity for ~37.36% of the population and prolongs the outbreak period from ~30 days to ~70 days, also aligning with the real case. We consider that this study lays the groundwork for the development of a new class of models and new insights for modelling heterogeneity.

PMID:39348420 | DOI:10.1371/journal.pcbi.1012497

PLoS Comput Biol. 2024 Sep 30;20(9):e1012489. doi: 10.1371/journal.pcbi.1012489. Online ahead of print.

ABSTRACT

Deep learning methods, trained on the increasing set of available protein 3D structures and sequences, have substantially impacted the protein modeling and design field. These advancements have facilitated the creation of novel proteins, or the optimization of existing ones designed for specific functions, such as binding a target protein. Despite the demonstrated potential of such approaches in designing general protein binders, their application in designing immunotherapeutics remains relatively underexplored. A relevant application is the design of T cell receptors (TCRs). Given the crucial role of T cells in mediating immune responses, redirecting these cells to tumor or infected target cells through the engineering of TCRs has shown promising results in treating diseases, especially cancer. However, the computational design of TCR interactions presents challenges for current physics-based methods, particularly due to the unique natural characteristics of these interfaces, such as low affinity and cross-reactivity. For this reason, in this study, we explored the potential of two structure-based deep learning protein design methods, ProteinMPNN and ESM-IF1, in designing fixed-backbone TCRs for binding target antigenic peptides presented by the MHC through different design scenarios. To evaluate TCR designs, we employed a comprehensive set of sequence- and structure-based metrics, highlighting the benefits of these methods in comparison to classical physics-based design methods and identifying deficiencies for improvement.

PMID:39348412 | DOI:10.1371/journal.pcbi.1012489

PLoS One. 2024 Sep 30;19(9):e0307083. doi: 10.1371/journal.pone.0307083. eCollection 2024.

ABSTRACT

Recent advances in deep learning and Vision-Language Models (VLM) have enabled efficient transfer to downstream tasks even when limited labelled training data is available, as well as for text to be directly compared to image content. These properties of VLMs enable new opportunities for the annotation and analysis of images. We test the potential of VLMs for landscape scenicness prediction, i.e., the aesthetic quality of a landscape, using zero- and few-shot methods. We experiment with few-shot learning by fine-tuning a single linear layer on a pre-trained VLM representation. We find that a model fitted to just a few hundred samples performs favourably compared to a model trained on hundreds of thousands of examples in a fully supervised way. We also explore the zero-shot prediction potential of contrastive prompting using positive and negative landscape aesthetic concepts. Our results show that this method outperforms a linear probe with few-shot learning when using a small number of samples to tune the prompt configuration. We introduce Landscape Prompt Ensembling (LPE), which is an annotation method for acquiring landscape scenicness ratings through rated text descriptions without needing an image dataset during annotation. We demonstrate that LPE can provide landscape scenicness assessments that are concordant with a dataset of image ratings. The success of zero- and few-shot methods combined with their ability to use text-based annotations highlights the potential for VLMs to provide efficient landscape scenicness assessments with greater flexibility.

PMID:39348404 | DOI:10.1371/journal.pone.0307083

PLOS Digit Health. 2024 Sep 30;3(9):e0000612. doi: 10.1371/journal.pdig.0000612. eCollection 2024 Sep.

ABSTRACT

Immigrants and refugees seeking admission to the United States must first undergo an overseas medical exam, overseen by the US Centers for Disease Control and Prevention (CDC), during which all persons ≥15 years old receive a chest x-ray to look for signs of tuberculosis. Although individual screening sites often implement quality control (QC) programs to ensure radiographs are interpreted correctly, the CDC does not currently have a method for conducting similar QC reviews at scale. We obtained digitized chest radiographs collected as part of the overseas immigration medical exam. Using radiographs from applicants 15 years old and older, we trained deep learning models to perform three tasks: identifying abnormal radiographs; identifying abnormal radiographs suggestive of tuberculosis; and identifying the specific findings (e.g., cavities or infiltrates) in abnormal radiographs. We then evaluated the models on both internal and external testing datasets, focusing on two classes of performance metrics: individual-level metrics, like sensitivity and specificity, and sample-level metrics, like accuracy in predicting the prevalence of abnormal radiographs. A total of 152,012 images (one image per applicant; mean applicant age 39 years) were used for model training. On our internal test dataset, our models performed well both in identifying abnormalities suggestive of TB (area under the curve [AUC] of 0.97; 95% confidence interval [CI]: 0.95, 0.98) and in estimating sample-level counts of the same (-2% absolute percentage error; 95% CIC: -8%, 6%). On the external test datasets, our models performed similarly well in identifying both generic abnormalities (AUCs ranging from 0.89 to 0.92) and those suggestive of TB (AUCs from 0.94 to 0.99). This performance was consistent across metrics, including those based on thresholded class predictions, like sensitivity, specificity, and F1 score. Strong performance relative to high-quality radiological reference standards across a variety of datasets suggests our models may make reliable tools for supporting chest radiography QC activities at CDC.

PMID:39348377 | DOI:10.1371/journal.pdig.0000612

IEEE J Biomed Health Inform. 2024 Sep 30;PP. doi: 10.1109/JBHI.2024.3471510. Online ahead of print.

ABSTRACT

Premature ventricular complexes (PVCs) are irregularities in heart rhythm where the ventricles contract earlier than expected, disrupting the normal cardiac cycle. Identifying the origin of PVCs before surgery is crucial as it can reduce operation duration, lower radiation exposure, and potentially enhance ablation success rates. Current detection methods face limitations in accuracy and data processing, often requiring large datasets and complex interpretations. This study presents PVCsNet, a deep-learning network specifically designed for classifying premature ventricular complexes (PVCs) in ECG images. It incorporates residual structures and attention mechanisms to enhance classification performance. PVCsNet consists of four 3×3 convolutional layers as feature extractors, followed by residual connections and attention blocks. This design enables the network to map image features to class probability distributions, enhancing performance even with limited data. Our experimental results demonstrate that using the SE Block with MaxPool and a ratio of 4, PVCsNet achieves an overall accuracy of 94.49%, with high precision in critical categories and a moderate parameter size. We successfully categorize the data into six distinct classes based on their origin locations in the heart: right ventricular outflow tract (RVOT), left ventricular outflow tract (LVOT), papillary muscle (PM), valvular annulus (VA), summit, and His-Purkinje system (HPS). Among these, RVOT is the most common and crucial origin of PVCs. PM and HPS are also significant origins due to their clinical implications. This study demonstrates the potential of PVCsNet in clinical diagnostics, providing promising results in classifying ECG images and contributing to future medical research and diagnosis.

PMID:39348245 | DOI:10.1109/JBHI.2024.3471510

Med Phys. 2024 Sep 30. doi: 10.1002/mp.17413. Online ahead of print.

ABSTRACT

BACKGROUND: Deep learning-based methods led to significant advancements in many areas of medical imaging, most of which are concerned with the reduction of artifacts caused by motion, scatter, or noise. However, with most neural networks being black boxes, they remain notoriously difficult to interpret, hindering their clinical implementation. In particular, it has been shown that networks exhibit invariances w.r.t. input features, that is, they learn to ignore certain information in the input data.

PURPOSE: To improve the interpretability of deep learning-based low-dose CT image denoising networks.

METHODS: We learn a complete data representation of low-dose input images using a conditional variational autoencoder (cVAE). In this representation, invariances of any given denoising network are then disentangled from the information it is not invariant to using a conditional invertible neural network (cINN). At test time, image-space invariances are generated by applying the inverse of the cINN and subsequent decoding using the cVAE. We propose two methods to analyze sampled invariances and to find those that correspond to alterations of anatomical structures.

RESULTS: The proposed method is applied to four popular deep learning-based low-dose CT image denoising networks. We find that the networks are not only invariant to noise amplitude and realizations, but also to anatomical structures.

CONCLUSIONS: The proposed method is capable of reconstructing and analyzing invariances of deep learning-based low-dose CT image denoising networks. This is an important step toward interpreting deep learning-based methods for medical imaging, which is essential for their clinical implementation.

PMID:39348044 | DOI:10.1002/mp.17413

Basic Res Cardiol. 2024 Sep 30. doi: 10.1007/s00395-024-01081-x. Online ahead of print.

ABSTRACT

Infarct size (IS) is the most robust end point for evaluating the success of preclinical studies on cardioprotection. The gold standard for IS quantification in ischemia/reperfusion (I/R) experiments is triphenyl tetrazolium chloride (TTC) staining, typically done manually. This study aimed to determine if automation through deep learning segmentation is a time-saving and valid alternative to standard IS quantification. High-resolution images from TTC-stained, macroscopic heart slices were retrospectively collected from pig experiments (n = 390) with I/R without/with cardioprotection to cover a wide IS range. Existing IS data from pig experiments, quantified using a standard method of manual and subsequent digital labeling of film-scan annotations, were used as reference. To automate the evaluation process with the aim to be more objective and save time, a deep learning pipeline was implemented; the collected images (n = 3869) were pre-processed by cropping and labeled (image annotations). To ensure their usability as training data for a deep learning segmentation model, IS was quantified from image annotations and compared to IS quantified using the existing film-scan annotations. A supervised deep learning segmentation model based on dynamic U-Net architecture was developed and trained. The evaluation of the trained model was performed by fivefold cross-validation (n = 220 experiments) and testing on an independent test set (n = 170 experiments). Performance metrics (Dice similarity coefficient [DSC], pixel accuracy [ACC], average precision [mAP]) were calculated. IS was then quantified from predictions and compared to IS quantified from image annotations (linear regression, Pearson's r; analysis of covariance; Bland-Altman plots). Performance metrics near 1 indicated a strong model performance on cross-validated data (DSC: 0.90, ACC: 0.98, mAP: 0.90) and on the test set data (DSC: 0.89, ACC: 0.98, mAP: 0.93). IS quantified from predictions correlated well with IS quantified from image annotations in all data sets (cross-validation: r = 0.98; test data set: r = 0.95) and analysis of covariance identified no significant differences. The model reduced the IS quantification time per experiment from approximately 90 min to 20 s. The model was further tested on a preliminary test set from experiments in isolated, saline-perfused rat hearts with regional I/R without/with cardioprotection (n = 27). There was also no significant difference in IS between image annotations and predictions, but the performance on the test set data from rat hearts was lower (DSC: 0.66, ACC: 0.91, mAP: 0.65). IS quantification using a deep learning segmentation model is a valid and time-efficient alternative to manual and subsequent digital labeling.

PMID:39348000 | DOI:10.1007/s00395-024-01081-x

Abdom Radiol (NY). 2024 Sep 30. doi: 10.1007/s00261-024-04607-y. Online ahead of print.

NO ABSTRACT

PMID:39347977 | DOI:10.1007/s00261-024-04607-y

J Cell Mol Med. 2024 Oct;28(19):e18591. doi: 10.1111/jcmm.18591.

ABSTRACT

The unique non-coding RNA molecule known as circular RNA (circRNA) is distinguished from conventional linear RNA by having a longer half-life, greater degree of conservation and inherent solidity. Extensive research has demonstrated the profound impact of circRNA expression on cellular drug sensitivity and therapeutic efficacy. There is an immediate need for the creation of efficient computational techniques to anticipate the potential correlations between circRNA and drug sensitivity, as classical biological research approaches are time-consuming and costly. In this work, we introduce a novel deep learning model called SNMGCDA, which aims to forecast the relationships between circRNA and drug sensitivity. SNMGCDA incorporates a diverse range of similarity networks, enabling the derivation of feature vectors for circRNAs and drugs using three distinct calculation methods. First, we utilize a sparse autoencoder for the extraction of drug characteristics. Subsequently, the application of non-negative matrix factorization (NMF) enables the identification of relationships between circRNAs and drugs based on their shared features. Additionally, the multi-head graph attention network is employed to capture the characteristics of circRNAs. After acquiring the characteristics from these three separate components, we combine them to form a unified and inclusive feature vector for each cluster of circRNA and drug. Finally, the relevant feature vectors and labels are inputted into a multilayer perceptron (MLP) to make predictions. The outcomes of the experiment, obtained through 5-fold cross-validation (5-fold CV) and 10-fold cross-validation (10-fold CV), demonstrate SNMGCDA outperforms five other state-of-art methods in terms of performance. Additionally, the majority of case studies have predominantly confirmed newly discovered correlations by SNMGCDA, thereby emphasizing its reliability in predicting potential relationships between circRNAs and drugs.

PMID:39347936 | DOI:10.1111/jcmm.18591

Eur Arch Otorhinolaryngol. 2024 Sep 30. doi: 10.1007/s00405-024-08979-1. Online ahead of print.

ABSTRACT

OBJECTIVES: This study's objectives are (1) to investigate the registration accuracy from intraoperative ultrasound (US) to histopathological images, (2) to assess the agreement and correlation between measurements in registered 3D US and histopathology, and (3) to train a nnUNet model for automatic segmentation of 3D US volumes of resected tongue specimens.

METHODS: Ten 3D US volumes were acquired, including the corresponding digitalized histopathological images (n = 29). Based on corresponding landmarks, the registrations between 3D US and histopathology images were calculated and evaluated using the target registration error (TRE). Tumor thickness and resection margins were measured based on three annotations: (1) manual histopathological tumor annotation (HTA), manual 3D US tumor annotation, and (2) the HTA registered in the 3D US. The agreement and correlation were computed between the measurements based on the HTA and those based on the manual US and registered HTA in US. A deep-learning model with nnUNet was trained on 151 3D US volumes. Segmentation metrics quantified the model's performance.

RESULTS: The median TRE was 0.42 mm. The smallest mean difference was between registered HTA in US and histopathology with 2.16 mm (95% CI - 1.31; 5.63) and a correlation of 0.924 (p < 0.001). The nnUNet predicted the tumor with a Dice similarity coefficient of 0.621, an average surface distance of 1.15 mm, and a Hausdorff distance of 3.70 mm.

CONCLUSION: Multimodal image registration enabled the HTA's registration in the US images and improved the agreement and correlation between the modalities. In the future, this could be used to annotate ground truth labels accurately.

PMID:39347853 | DOI:10.1007/s00405-024-08979-1

J Cardiovasc Med (Hagerstown). 2024 Nov 1;25(11):781-788. doi: 10.2459/JCM.0000000000001670. Epub 2024 Sep 12.

ABSTRACT

BACKGROUND: Cardiovascular risk assessment is a critical component of healthcare, guiding preventive and therapeutic strategies. In this study, we developed and evaluated an image-based electrocardiogram (ECG) analyzing an artificial intelligence (AI) model that estimates biological age and mortality risk.

METHODS: Using a dataset of 978 319 ECGs from 250 145 patients at Seoul National University Bundang Hospital, we developed a deep-learning model utilizing printed 12-lead ECG images to estimate patients' age (ECG-Age) and 1- and 5-year mortality risks. The model was validated externally using the CODE-15% dataset from Brazil.

RESULTS: The ECG-Age showed a high correlation with chronological age in both the internal and external validation datasets (Pearson's R = 0.888 and 0.852, respectively). In the internal validation, the direct mortality risk prediction models showed area under the curves (AUCs) of 0.843 and 0.867 for 5- and 1-year all-cause mortality, respectively. For 5- and 1-year cardiovascular mortality, the AUCs were 0.920 and 0.916, respectively. In the CODE-15%, the mortality risk predictions showed AUCs of 0.818 and 0.836 for the prediction of 5- and 1-year all-cause mortality, respectively. Compared to the neutral Delta-Age (ECG-Age - chronological age) group, hazard ratios for deaths were 1.88 [95% confidence interval (CI): 1.14-3.92], 2.12 (95% CI: 1.15-3.92), 4.46 (95% CI: 2.22-8.96) and 7.68 (95% CI: 3.32-17.76) for positive Delta-Age groups (5-10, 10-15, 15-20, >20), respectively.

CONCLUSION: An image-based AI-ECG model is a feasible tool for estimating biological age and assessing all-cause and cardiovascular mortality risks, providing a practical approach for utilizing standardized ECG images in predicting long-term health outcomes.

PMID:39347726 | DOI:10.2459/JCM.0000000000001670

Photoacoustics. 2024 Sep 11;40:100645. doi: 10.1016/j.pacs.2024.100645. eCollection 2024 Dec.

ABSTRACT

Deep learning has been used to improve photoacoustic (PA) image reconstruction. One major challenge is that errors cannot be quantified to validate predictions when ground truth is unknown. Validation is key to quantitative applications, especially using limited-bandwidth ultrasonic linear detector arrays. Here, we propose a hybrid Bayesian convolutional neural network (Hybrid-BCNN) to jointly predict PA image and segmentation with error (uncertainty) predictions. Each output pixel represents a probability distribution where error can be quantified. The Hybrid-BCNN was trained with simulated PA data and applied to both simulations and experiments. Due to the sparsity of PA images, segmentation focuses Hybrid-BCNN on minimizing the loss function in regions with PA signals for better predictions. The results show that accurate PA segmentations and images are obtained, and error predictions are highly statistically correlated to actual errors. To leverage error predictions, confidence processing created PA images above a specific confidence level.

PMID:39347464 | PMC:PMC11424948 | DOI:10.1016/j.pacs.2024.100645

J Dent Sci. 2024 Oct;19(4):2140-2149. doi: 10.1016/j.jds.2024.03.017. Epub 2024 Mar 27.

ABSTRACT

BACKGROUND/PURPOSE: The All-on-4® treatment concept is a dental procedure that utilizes only four dental implants to support a fixed prosthesis, providing full-arch rehabilitation with affordable cost and speedy treatment courses. Although the placement of all-on-4® implants has been researched in the past, little attention was paid to the structural design of the prosthetic framework.

MATERIALS AND METHODS: This research proposed a new approach to optimize the structure of denture framework called BESO-Net, which is a bidirectional evolutionary structural optimization (BESO) based convolutional neural network (CNN). The approach aimed to reduce the use of material for the framework, such as Ti-6Al-4V, while maintaining structural strength. The BESO-Net was designed as a one-dimensional CNN based on Inception V3, trained using finite element analysis (FEA) data from 14,994 design configurations, and evaluated its training performance, generalization capability, and computation efficiency.

RESULTS: The results suggested that BESO-Net accurately predicted the optimal structure of the denture framework in various mandibles with different implant and load settings. The average error was found to be 0.29% for compliance and 11.26% for shape error when compared to the traditional BESO combined with FEA. Additionally, the computational time required for structural optimization was significantly reduced from 6.5 h to 45 s.

CONCLUSION: The proposed approach demonstrates its applicability in clinical settings to quickly find personalized All-on-4® framework structure that can significantly reduce material consumption while maintaining sufficient stiffness.

PMID:39347035 | PMC:PMC11437609 | DOI:10.1016/j.jds.2024.03.017

Biomed Opt Express. 2024 Jul 3;15(8):4438-4452. doi: 10.1364/BOE.524946. eCollection 2024 Aug 1.

ABSTRACT

Coronary artery calcification (CAC) is a marker of atherosclerosis and is thought to be associated with worse clinical outcomes. However, evidence from large-scale high-resolution imaging data is lacking. We proposed a novel deep learning method that can automatically identify and quantify CAC in massive intravascular OCT data trained using efficiently generated sparse labels. 1,106,291 OCT images from 1,048 patients were collected and utilized to train and evaluate the method. The Dice similarity coefficient for CAC segmentation and the accuracy for CAC classification are 0.693 and 0.932, respectively, close to human-level performance. Applying the method to 1259 ST-segment elevated myocardial infarction patients imaged with OCT, we found that patients with a greater extent and more severe calcification in the culprit vessels were significantly more likely to have major adverse cardiovascular and cerebrovascular events (MACCE) (p < 0.05), while the CAC in non-culprit vessels did not differ significantly between MACCE and non-MACCE groups.

PMID:39347010 | PMC:PMC11427185 | DOI:10.1364/BOE.524946