J Am Med Inform Assoc. 2024 Dec 28:ocae316. doi: 10.1093/jamia/ocae316. Online ahead of print.

ABSTRACT

OBJECTIVE: Early detection of surgical complications allows for timely therapy and proactive risk mitigation. Machine learning (ML) can be leveraged to identify and predict patient risks for postoperative complications. We developed and validated the effectiveness of predicting postoperative complications using a novel surgical Variational Autoencoder (surgVAE) that uncovers intrinsic patterns via cross-task and cross-cohort presentation learning.

MATERIALS AND METHODS: This retrospective cohort study used data from the electronic health records of adult surgical patients over 4 years (2018-2021). Six key postoperative complications for cardiac surgery were assessed: acute kidney injury, atrial fibrillation, cardiac arrest, deep vein thrombosis or pulmonary embolism, blood transfusion, and other intraoperative cardiac events. We compared surgVAE's prediction performance against widely-used ML models and advanced representation learning and generative models under 5-fold cross-validation.

RESULTS: 89 246 surgeries (49% male, median [IQR] age: 57 [45-69]) were included, with 6502 in the targeted cardiac surgery cohort (61% male, median [IQR] age: 60 [53-70]). surgVAE demonstrated generally superior performance over existing ML solutions across postoperative complications of cardiac surgery patients, achieving macro-averaged AUPRC of 0.409 and macro-averaged AUROC of 0.831, which were 3.4% and 3.7% higher, respectively, than the best alternative method (by AUPRC scores). Model interpretation using Integrated Gradients highlighted key risk factors based on preoperative variable importance.

DISCUSSION AND CONCLUSION: Our advanced representation learning framework surgVAE showed excellent discriminatory performance for predicting postoperative complications and addressing the challenges of data complexity, small cohort sizes, and low-frequency positive events. surgVAE enables data-driven predictions of patient risks and prognosis while enhancing the interpretability of patient risk profiles.

PMID:39731515 | DOI:10.1093/jamia/ocae316

Adv Sci (Weinh). 2024 Dec 27:e2412670. doi: 10.1002/advs.202412670. Online ahead of print.

ABSTRACT

The biobased production of chemicals is essential for advancing a sustainable chemical industry. 1,5-Pentanediol (1,5-PDO), a five-carbon diol with considerable industrial relevance, has shown limited microbial production efficiency until now. This study presents the development and optimization of a microbial system to produce 1,5-PDO from glucose in Corynebacterium glutamicum via the l-lysine-derived pathway. Engineering began with creating a strain capable of producing 5-hydroxyvaleric acid (5-HV), a key precursor to 1,5-PDO, by incorporating enzymes from Pseudomonas putida (DavB, DavA, and DavT) and Escherichia coli (YahK). Two conversion pathways for further converting 5-HV to 1,5-PDO are evaluated, with the CoA-independent pathway-utilizing Mycobacterium marinum carboxylic acid reductase (CAR) and E. coli YqhD-proving greater efficiency. Further optimization continues with chromosomal integration of the 5-HV module, increasing 1,5-PDO production to 5.48 g L-1. An additional screening of 13 CARs identifies Mycobacterium avium K-10 (MAP1040) as the most effective, and its engineered M296E mutant further increases production to 23.5 g L-1. A deep-learning analysis reveals that Gluconobacter oxydans GOX1801 resolves the limitations of NADPH, allowing the final strain to produce 43.4 g L-1 1,5-PDO without 5-HV accumulation in fed-batch fermentation. This study demonstrates systematic approaches to optimizing microbial biosynthesis, positioning C. glutamicum as a promising platform for sustainable 1,5-PDO production.

PMID:39731342 | DOI:10.1002/advs.202412670

Adv Sci (Weinh). 2024 Dec 27:e2406797. doi: 10.1002/advs.202406797. Online ahead of print.

ABSTRACT

Digital PCR (dPCR) has transformed nucleic acid diagnostics by enabling the absolute quantification of rare mutations and target sequences. However, traditional dPCR detection methods, such as those involving flow cytometry and fluorescence imaging, may face challenges due to high costs, complexity, limited accuracy, and slow processing speeds. In this study, SAM-dPCR is introduced, a training-free open-source bioanalysis paradigm that offers swift and precise absolute quantification of biological samples. SAM-dPCR leverages the robustness of the zero-shot Segment Anything Model (SAM) to achieve rapid processing times (<4 seconds) with an accuracy exceeding 97.10%. This method has been extensively validated across diverse samples and reactor morphologies, demonstrating its broad applicability. Utilizing standard laboratory fluorescence microscopes, SAM-dPCR can measure nucleic acid template concentrations ranging from 0.154 copies µL-1 to 1.295 × 103 copies µL-1 for droplet dPCR and 0.160 × 103 to 3.629 × 103 copies µL-1 for microwell dPCR. Experimental validation shows a strong linear relationship (r2 > 0.96) between expected and determined sample concentrations. SAM-dPCR offers high accuracy, accessibility, and the ability to address bioanalytical needs in resource-limited settings, as it does not rely on hand-crafted "ground truth" data.

PMID:39731324 | DOI:10.1002/advs.202406797

J Med Eng Technol. 2024 Dec 27:1-13. doi: 10.1080/03091902.2024.2438150. Online ahead of print.

ABSTRACT

Nowadays, photoplethysmograph (PPG) technology is being used more often in smart devices and mobile phones due to advancements in information and communication technology in the health field, particularly in monitoring cardiac activities. Developing generative models to generate synthetic PPG signals requires overcoming challenges like data diversity and limited data available for training deep learning models. This paper proposes a generative model by adopting a genetic programming (GP) approach to generate increasingly diversified and accurate data using an initial PPG signal sample. Unlike conventional regression, the GP approach automatically determines the structure and combinations of a mathematical model. Given that mean square error (MSE) of 0.0001, root mean square error (RMSE) of 0.01, and correlation coefficient of 0.999, the proposed approach outperformed other approaches and proved effective in terms of efficiency and applicability in resource-constrained environments.

PMID:39731227 | DOI:10.1080/03091902.2024.2438150

Acta Ophthalmol. 2024 Dec 27. doi: 10.1111/aos.17434. Online ahead of print.

ABSTRACT

PURPOSE: The relationship between retinal morphology, as assessed by optical coherence tomography (OCT), and retinal function in microperimetry (MP) has not been well studied, despite its increasing importance as an essential functional endpoint for clinical trials and emerging therapies in retinal diseases. Normative databases of healthy ageing eyes are largely missing from literature.

METHODS: Healthy subjects above 50 years were examined using two MP devices, MP-3 (NIDEK) and MAIA (iCare). An identical grid, encompassing 45 stimuli was used for retinal sensitivity (RS) assessment. Deep-learning-based algorithms performed automated segmentation of ellipsoid zone (EZ), outer nuclear layer (ONL), ganglion cell layer (GCL) and retinal nerve fibre layer (RNFL) from OCT volumes (Spectralis, Heidelberg). Pointwise co-registration between each MP stimulus and corresponding location on OCT was performed via registration algorithm. Effect of age, eccentricity and layer thickness on RS was assessed by mixed effect models.

RESULTS: Three thousand six hundred stimuli from twenty eyes of twenty patients were included. Mean patient age was 68.9 ± 10.9 years. Mean RS for the first exam was 28.65 ± 2.49 dB and 25.5 ± 2.81 dB for MP-3 and MAIA, respectively. Increased EZ thickness, ONL thickness and GCL thickness were significantly correlated with increased RS (all p < 0.001). Univariate models showed lower RS with advanced age and higher eccentricity (both p < 0.05).

CONCLUSION: This work provides reference values for healthy age-related EZ and ONL-thicknesses without impairment of visual function and evidence for RS decrease with eccentricity and increasing age. This information is crucial for interpretation of future clinical trials in disease.

PMID:39731225 | DOI:10.1111/aos.17434

J Cheminform. 2024 Dec 27;16(1):146. doi: 10.1186/s13321-024-00941-x.

ABSTRACT

Naming chemical compounds systematically is a complex task governed by a set of rules established by the International Union of Pure and Applied Chemistry (IUPAC). These rules are universal and widely accepted by chemists worldwide, but their complexity makes it challenging for individuals to consistently apply them accurately. A translation method can be employed to address this challenge. Accurate translation of chemical compounds from SMILES notation into their corresponding IUPAC names is crucial, as it can significantly streamline the laborious process of naming chemical structures. Here, we present STOUT (SMILES-TO-IUPAC-name translator) V2, which addresses this challenge by introducing a transformer-based model that translates string representations of chemical structures into IUPAC names. Trained on a dataset of nearly 1 billion SMILES strings and their corresponding IUPAC names, STOUT V2 demonstrates exceptional accuracy in generating IUPAC names, even for complex chemical structures. The model's ability to capture intricate patterns and relationships within chemical structures enables it to generate precise and standardised IUPAC names. While established deterministic algorithms remain the gold standard for systematic chemical naming, our work, enabled by access to OpenEye's Lexichem software through an academic license, demonstrates the potential of neural approaches to complement existing tools in chemical nomenclature.Scientific contribution STOUT V2, built upon transformer-based models, is a significant advancement from our previous work. The web application enhances its accessibility and utility. By making the model and source code fully open and well-documented, we aim to promote unrestricted use and encourage further development.

PMID:39731139 | DOI:10.1186/s13321-024-00941-x

BMC Cancer. 2024 Dec 27;24(1):1581. doi: 10.1186/s12885-024-13320-4.

ABSTRACT

Glioblastoma Multiforme (GBM), classified as a grade IV glioma by the World Health Organization (WHO), is a prevalent and notably aggressive form of brain tumor derived from glial cells. It stands as one of the most severe forms of primary brain cancer in humans. The median survival time of GBM patients is only 12-15 months, making it the most lethal type of brain tumor. Every year, about 200,000 people worldwide succumb to this disease. GBM is also highly heterogeneous, meaning that its characteristics and behavior vary widely among different patients. This leads to different outcomes and survival times for each individual. Predicting the survival of GBM patients accurately can have multiple benefits. It can enable optimal and personalized treatment planning based on the patient's condition and prognosis. It can also support the patients and their families to cope with the possible outcomes and make informed decisions about their care and quality of life. Furthermore, it can assist the researchers and scientists to discover the most relevant biomarkers, features, and mechanisms of the disease and to design more effective and personalized therapies. Artificial intelligence methods, such as machine learning and deep learning, have been widely applied to survival prediction in various fields, such as breast cancer, lung cancer, gastric cancer, cervical cancer, liver cancer, prostate cancer, and covid 19. This systematic review summarizes the current state-of-the-art methods for predicting glioblastoma survival using different types of input data, such as clinical features, molecular markers, imaging features, radiomics features, omics data or a combination of them. Following PRISMA guidelines, we searched databases from 2015 to 2024, reviewing 107 articles meeting our criteria. We analyzed the data sources, methods, performance metrics and outcomes of the studies. We found that random forest was the most popular method, and a combination of radiomics and clinical data was the most common input data.

PMID:39731064 | DOI:10.1186/s12885-024-13320-4

BMC Bioinformatics. 2024 Dec 27;25(1):392. doi: 10.1186/s12859-024-06013-z.

ABSTRACT

BACKGROUND: Cancer classification has consistently been a challenging problem, with the main difficulties being high-dimensional data and the collection of patient samples. Concretely, obtaining patient samples is a costly and resource-intensive process, and imbalances often exist between samples. Moreover, expression data is characterized by high dimensionality, small samples and high noise, which could easily lead to struggles such as dimensionality catastrophe and overfitting. Thus, we incorporate prior knowledge from the pathway and combine AutoEncoder and Generative Adversarial Network (GAN) to solve these difficulties.

RESULTS: In this study, we propose an effective and efficient deep learning method, named AEGAN, which combines the capabilities of AutoEncoder and GAN to generate synthetic samples of the minority class in imbalanced gene expression data. The proposed data balancing technique has been demonstrated to be useful for cancer classification and improving the performance of classifier models. Additionally, we integrate prior knowledge from the pathway and employ the pathifier algorithm to calculate pathway scores for each sample. This data augmentation approach, referred to as AEGAN-Pathifier, not only preserves the biological functionality of the data but also possesses dimensional reduction capabilities. Through validation with various classifiers, the experimental results show an improvement in classifier performance.

CONCLUSION: AEGAN-Pathifier shows improved performance on the imbalanced datasets GSE25066, GSE20194, BRCA and Liver24. Results from various classifiers indicate that AEGAN-Pathifier has good generalization capability.

PMID:39731019 | DOI:10.1186/s12859-024-06013-z

BMC Genomics. 2024 Dec 27;25(1):1252. doi: 10.1186/s12864-024-11173-6.

ABSTRACT

BACKGROUND: The subcellular localization of mRNA plays a crucial role in gene expression regulation and various cellular processes. However, existing wet lab techniques like RNA-FISH are usually time-consuming, labor-intensive, and limited to specific tissue types. Researchers have developed several computational methods to predict mRNA subcellular localization to address this. These methods face the problem of class imbalance in multi-label classification, causing models to favor majority classes and overlook minority classes during training. Additionally, traditional feature extraction methods have high computational costs, incomplete features, and may lead to the loss of critical information. On the other hand, deep learning methods face challenges related to hardware performance and training time when handling complex sequences. They may suffer from the curse of dimensionality and overfitting problems. Therefore, there is an urgent need for more efficient and accurate prediction models.

RESULTS: To address these issues, we propose a multi-label classifier, EDCLoc, for predicting mRNA subcellular localization. EDCLoc reduces training pressure through a stepwise pooling strategy and applies grouped convolution blocks of varying sizes at different levels, combined with residual connections, to achieve efficient feature extraction and gradient propagation. The model employs global max pooling at the end to further reduce feature dimensions and highlight key features. To tackle class imbalance, we improved the focal loss function to enhance the model's focus on minority classes. Evaluation results show that EDCLoc outperforms existing methods in most subcellular regions. Additionally, the position weight matrix extracted by multi-scale CNN filters can match known RNA-binding protein motifs, demonstrating EDCLoc's effectiveness in capturing key sequence features.

CONCLUSIONS: EDCLoc outperforms existing prediction tools in most subcellular regions and effectively mitigates class imbalance issues in multi-label classification. These advantages make EDCLoc a reliable choice for multi-label mRNA subcellular localization. The dataset and source code used in this study are available at https://github.com/DellCode233/EDCLoc .

PMID:39731012 | DOI:10.1186/s12864-024-11173-6

Sci Rep. 2024 Dec 28;14(1):31193. doi: 10.1038/s41598-024-82562-w.

ABSTRACT

Deep Convolutional Neural Networks (DCNNs), due to their high computational and memory requirements, face significant challenges in deployment on resource-constrained devices. Network Pruning, an essential model compression technique, contributes to enabling the efficient deployment of DCNNs on such devices. Compared to traditional rule-based pruning methods, Reinforcement Learning(RL)-based automatic pruning often yields more effective pruning strategies through its ability to learn and adapt. However, the current research only set a single agent to explore the optimal pruning rate for all convolutional layers, ignoring the interactions and effects among multiple layers. To address this challenge, this paper proposes an automatic Filter Pruning method with a multi-agent reinforcement learning algorithm QMIX, named QMIX_FP. The multi-layer structure of DCNNs is modeled as a multi-agent system, which considers the varying sensitivity of each convolutional layer to the entire DCNN and the interactions among them. We employ the multi-agent reinforcement learning algorithm QMIX, where individual agent contributes to the system monotonically, to explore the optimal iterative pruning strategy for each convolutional layer. Furthermore, fine-tuning the pruned network using knowledge distillation accelerates model performance improvement. The efficiency of this method is demonstrated on two benchmark DCNNs, including VGG-16 and AlexNet, over CIFAR-10 and CIFAR-100 datasets. Extensive experiments under different scenarios show that QMIX_FP not only reduces the computational and memory requirements of the networks but also maintains their accuracy, making it a significant advancement in the field of model compression and efficient deployment of deep learning models on resource-constrained devices.

PMID:39730902 | DOI:10.1038/s41598-024-82562-w

Commun Biol. 2024 Dec 27;7(1):1705. doi: 10.1038/s42003-024-07411-y.

ABSTRACT

Understanding the function of proteins is of great significance for revealing disease pathogenesis and discovering new targets. Benefiting from the explosive growth of the protein universal, deep learning has been applied to accelerate the protein annotation cycle from different biological modalities. However, most existing deep learning-based methods not only fail to effectively fuse different biological modalities, resulting in low-quality protein representations, but also suffer from the convergence of suboptimal solution caused by sparse label representations. Aiming at the above issue, we propose a multiprocedural approach for fusing heterogeneous biological modalities and annotating protein functions, i.e., MIF2GO (Multimodal Information Fusion to infer Gene Ontology terms), which sequentially fuses up to six biological modalities ranging from different biological levels in three steps, thus leading to powerful protein representations. Evaluation results on seven benchmark datasets show that the proposed method not only considerably outperforms state-of-the-art performance, but also demonstrates great robustness and generalizability across species. Besides, we also present biological insights into the associations between those modalities and protein functions. This research provides a robust framework for integrating multimodal biological data, offering a scalable solution for protein function annotation, ultimately facilitating advancements in precision medicine and the discovery of novel therapeutic strategies.

PMID:39730886 | DOI:10.1038/s42003-024-07411-y

Invest Radiol. 2024 Oct 10. doi: 10.1097/RLI.0000000000001128. Online ahead of print.

ABSTRACT

OBJECTIVES: The aim of this study was to evaluate the feasibility and efficacy of visual scoring, low-attenuation volume (LAV), and deep learning methods for estimating emphysema extent in x-ray dose photon-counting detector computed tomography (PCD-CT), aiming to explore future dose reduction potentials.

METHODS: One hundred one prospectively enrolled patients underwent noncontrast low- and chest x-ray dose CT scans in the same study using PCD-CT. Overall image quality, sharpness, and noise, as well as visual emphysema pattern (no, trace, mild, moderate, confluent, and advanced destructive emphysema; as defined by the Fleischner Society), were independently assessed by 2 experienced radiologists for low- and x-ray dose images, followed by an expert consensus read. In the second step, automated emphysema quantification was performed using an established LAV algorithm with a threshold of -950 HU and a commercially available deep learning model for automated emphysema quantification. Automated estimations of emphysema extent were converted and compared with visual scoring ratings.

RESULTS: X-ray dose scans exhibited a significantly lower computed tomography dose index than low-dose scans (low-dose: 0.66 ± 0.16 mGy, x-ray dose: 0.11 ± 0.03 mGy, P < 0.001). Interreader agreement between low- and x-ray dose for visual emphysema scoring was excellent (κ = 0.83). Visual emphysema scoring consensus showed good agreement between low-dose and x-ray dose scans (κ = 0.70), with significant and strong correlation (Spearman ρ = 0.79). Although trace emphysema was underestimated in x-ray dose scans, there was no significant difference in the detection of higher-grade (mild to advanced destructive) emphysema (P = 0.125) between the 2 scan doses. Although predicted emphysema volumes on x-ray dose scans for the LAV method showed strong and the deep learning model excellent significant correlations with predictions on low-dose scans, both methods significantly overestimated emphysema volumes on lower quality scans (P < 0.001), with the deep learning model being more robust. Further, deep learning emphysema severity estimations showed higher agreement (κ = 0.65) and correlation (Spearman ρ = 0.64) with visual scoring for low-dose scans than LAV predictions (κ = 0.48, Spearman ρ = 0.45).

CONCLUSIONS: The severity of emphysema can be reliably estimated using visual scoring on CT scans performed with x-ray equivalent doses on a PCD-CT. A deep learning algorithm demonstrated good agreement and strong correlation with the visual scoring method on low-dose scans. However, both the deep learning and LAV algorithms overestimated emphysema extent on x-ray dose scans. Nonetheless, x-ray equivalent radiation dose scans may revolutionize the detection and monitoring of disease in chronic obstructive pulmonary disease patients.

PMID:39729642 | DOI:10.1097/RLI.0000000000001128

PLoS One. 2024 Dec 27;19(12):e0315245. doi: 10.1371/journal.pone.0315245. eCollection 2024.

ABSTRACT

The increasing utilization of deep learning models in drug repositioning has proven to be highly efficient and effective. In this study, we employed an integrated deep-learning model followed by traditional drug screening approach to screen a library of FDA-approved drugs, aiming to identify novel inhibitors targeting the TNF-α converting enzyme (TACE). TACE, also known as ADAM17, plays a crucial role in the inflammatory response by converting pro-TNF-α to its active soluble form and cleaving other inflammatory mediators, making it a promising target for therapeutic intervention in diseases such as rheumatoid arthritis. Reference datasets containing active and decoy compounds specific to TACE were obtained from the DUD-E database. Using RDKit, a cheminformatics toolkit, we extracted molecular features from these compounds. We applied the GraphConvMol model within the DeepChem framework, which utilizes graph convolutional networks, to build a predictive model based on the DUD-E datasets. Our trained model was subsequently used to predict the TACE inhibitory potential of FDA-approved drugs. From these predictions, Vorinostat was identified as a potential TACE inhibitor. Moreover, molecular docking and molecular dynamics simulation were conducted to validate these findings, using BMS-561392 as a reference TACE inhibitor. Vorinostat, originally an FDA-approved drug for cancer treatment, exhibited strong binding interactions with key TACE residues, suggesting its repurposing potential. Biological evaluation with RAW 264.7 cell confirmed the computational results, demonstrating that Vorinostat exhibited comparable inhibitory activity against TACE. In conclusion, our study highlights the capability of deep learning models to enhance virtual screening efforts in drug discovery, efficiently identifying potential candidates for specific targets such as TACE. Vorinostat, as a newly identified TACE inhibitor, holds promise for further exploration and investigation in the treatment of inflammatory diseases like rheumatoid arthritis.

PMID:39729480 | DOI:10.1371/journal.pone.0315245

PLoS One. 2024 Dec 27;19(12):e0311081. doi: 10.1371/journal.pone.0311081. eCollection 2024.

ABSTRACT

BACKGROUND: Dyspnoea is one of the emergency department's (ED) most common and deadly chief complaints, but frequently misdiagnosed and mistreated. We aimed to design a diagnostic decision support which classifies dyspnoeic ED visits into acute heart failure (AHF), exacerbation of chronic obstructive pulmonary disease (eCOPD), pneumonia and "other diagnoses" by using deep learning and complete, unselected data from an entire regional health care system.

METHODS: In this cross-sectional study, we included all dyspnoeic ED visits of patients ≥ 18 years of age at the two EDs in the region of Halland, Sweden, 07/01/2017-12/31/2019. Data from the complete regional health care system within five years prior to the ED visit were analysed. Gold standard diagnoses were defined as the subsequent in-hospital or ED discharge notes, and a subsample was manually reviewed by emergency medicine experts. A novel deep learning model, the clinical attention-based recurrent encoder network (CareNet), was developed. Cohort performance was compared to a simpler CatBoost model. A list of all variables and their importance for diagnosis was created. For each unique patient visit, the model selected the most important variables, analysed them and presented them to the clinician interpretably by taking event time and clinical context into account. AUROC, sensitivity and specificity were compared.

FINDINGS: The most prevalent diagnoses among the 10,315 dyspnoeic ED visits were AHF (15.5%), eCOPD (14.0%) and pneumonia (13.3%). Median number of unique events, i.e., registered clinical data with time stamps, per ED visit was 1,095 (IQR 459-2,310). CareNet median AUROC was 87.0%, substantially higher than the CatBoost model´s (81.4%). CareNet median sensitivity for AHF, eCOPD, and pneumonia was 74.5%, 92.6%, and 54.1%, respectively, with a specificity set above 75.0, slightly inferior to that of the CatBoost baseline model. The model assembled a list of 1,596 variables by importance for diagnosis, on top were prior diagnoses of heart failure or COPD, daily smoking, atrial fibrillation/flutter, life management difficulties and maternity care. Each patient visit received their own unique attention plot, graphically displaying important clinical events for the diagnosis.

INTERPRETATION: We designed a novel interpretable deep learning model for diagnosis in emergency department dyspnoea patients by analysing unselected data from a complete regional health care system.

PMID:39729465 | DOI:10.1371/journal.pone.0311081

Oral Radiol. 2024 Dec 27. doi: 10.1007/s11282-024-00799-7. Online ahead of print.

ABSTRACT

OBJECTIVE: The aim of this study is to determine the contact relationship and position of impacted mandibular third molar teeth (IMM3) with the mandibular canal (MC) in panoramic radiography (PR) images using deep learning (DL) models trained with the help of cone beam computed tomography (CBCT) and DL to compare the performances of the architectures.

METHODS: In this study, a total of 546 IMM3s from 290 patients with CBCT and PR images were included. The performances of SqueezeNet, GoogLeNet, and Inception-v3 architectures in solving four problems on two different regions of interest (RoI) were evaluated.

RESULTS: The SqueezeNet architecture performed the best on the vertical RoI, showing 93.2% accuracy in the identification of the 2nd problem (contact relationship buccal or lingual). Inception-v3 showed the highest performance with 84.8% accuracy in horizontal RoI for the 1st problem (contact relationship-no contact relationship), GoogLeNet showed 77.4% accuracy in horizontal RoI for the 4th problem (contact relationship buccal, lingual, other category, or no contact relationship), and GoogLeNet showed 70.0% accuracy in horizontal RoI for the 3rd problem (contact relationship buccal, lingual, or other category).

CONCLUSION: This study found that the Inception-v3 model showed the highest accuracy values in determining the contact relationship, and SqueezeNet architecture showed the highest accuracy values in determining the position of IMM3 relative to MC in the presence of a contact relationship.

PMID:39729224 | DOI:10.1007/s11282-024-00799-7

Tomography. 2024 Dec 23;10(12):2144-2157. doi: 10.3390/tomography10120151.

ABSTRACT

BACKGROUND/OBJECTIVES: Calculating the radiation dose from CT in 18F-PET/CT examinations poses a significant challenge. The objective of this study is to develop a deep learning-based automated program that standardizes the measurement of radiation doses.

METHODS: The torso CT was segmented into six distinct regions using TotalSegmentator. An automated program was employed to extract the necessary information and calculate the effective dose (ED) of PET/CT. The accuracy of our automated program was verified by comparing the EDs calculated by the program with those determined by a nuclear medicine physician (n = 30). Additionally, we compared the EDs obtained from an older PET/CT scanner with those from a newer PET/CT scanner (n = 42).

RESULTS: The CT ED calculated by the automated program was not significantly different from that calculated by the nuclear medicine physician (3.67 ± 0.61 mSv and 3.62 ± 0.60 mSv, respectively, p = 0.7623). Similarly, the total ED showed no significant difference between the two calculation methods (8.10 ± 1.40 mSv and 8.05 ± 1.39 mSv, respectively, p = 0.8957). A very strong correlation was observed in both the CT ED and total ED between the two measurements (r2 = 0.9981 and 0.9996, respectively). The automated program showed excellent repeatability and reproducibility. When comparing the older and newer PET/CT scanners, the PET ED was significantly lower in the newer scanner than in the older scanner (4.39 ± 0.91 mSv and 6.00 ± 1.17 mSv, respectively, p < 0.0001). Consequently, the total ED was significantly lower in the newer scanner than in the older scanner (8.22 ± 1.53 mSv and 9.65 ± 1.34 mSv, respectively, p < 0.0001).

CONCLUSIONS: We successfully developed an automated program for calculating the ED of torso 18F-PET/CT. By integrating a deep learning model, the program effectively eliminated inter-operator variability.

PMID:39728913 | DOI:10.3390/tomography10120151

Tomography. 2024 Dec 18;10(12):2073-2086. doi: 10.3390/tomography10120147.

ABSTRACT

OBJECTIVES: We evaluated the noise reduction effects of deep learning reconstruction (DLR) and hybrid iterative reconstruction (HIR) in brain computed tomography (CT).

METHODS: CT images of a 16 cm dosimetry phantom, a head phantom, and the brains of 11 patients were reconstructed using filtered backprojection (FBP) and various levels of DLR and HIR. The slice thickness was 5, 2.5, 1.25, and 0.625 mm. Phantom imaging was also conducted at various tube currents. The noise reduction ratio was calculated using FBP as the reference. For patient imaging, overall image quality was visually compared between DLR and HIR images that exhibited similar noise reduction ratios.

RESULTS: The noise reduction ratio increased with increasing levels of DLR and HIR in phantom and patient imaging. For DLR, noise reduction was more pronounced with decreasing slice thickness, while such thickness dependence was less evident for HIR. Although the noise reduction effects of DLR were similar between the head phantom and patients, they differed for the dosimetry phantom. Variations between imaging objects were small for HIR. The noise reduction ratio was low at low tube currents for the dosimetry phantom using DLR; otherwise, the influence of the tube current was small. In terms of visual image quality, DLR outperformed HIR in 1.25 mm thick images but not in thicker images.

CONCLUSIONS: The degree of noise reduction using DLR depends on the slice thickness, tube current, and imaging object in addition to the level of DLR, which should be considered in the clinical use of DLR. DLR may be particularly beneficial for thin-slice imaging.

PMID:39728909 | DOI:10.3390/tomography10120147

Tomography. 2024 Dec 12;10(12):2038-2057. doi: 10.3390/tomography10120145.

ABSTRACT

Background/Objectives: Breast cancer is a leading cause of mortality among women in Taiwan and globally. Non-invasive imaging methods, such as mammography and ultrasound, are critical for early detection, yet standalone modalities have limitations in regard to their diagnostic accuracy. This study aims to enhance breast cancer detection through a cross-modality fusion approach combining mammography and ultrasound imaging, using advanced convolutional neural network (CNN) architectures. Materials and Methods: Breast images were sourced from public datasets, including the RSNA, the PAS, and Kaggle, and categorized into malignant and benign groups. Data augmentation techniques were used to address imbalances in the ultrasound dataset. Three models were developed: (1) pre-trained CNNs integrated with machine learning classifiers, (2) transfer learning-based CNNs, and (3) a custom-designed 17-layer CNN for direct classification. The performance of the models was evaluated using metrics such as accuracy and the Kappa score. Results: The custom 17-layer CNN outperformed the other models, achieving an accuracy of 0.964 and a Kappa score of 0.927. The transfer learning model achieved moderate performance (accuracy 0.846, Kappa 0.694), while the pre-trained CNNs with machine learning classifiers yielded the lowest results (accuracy 0.780, Kappa 0.559). Cross-modality fusion proved effective in leveraging the complementary strengths of mammography and ultrasound imaging. Conclusions: This study demonstrates the potential of cross-modality imaging and tailored CNN architectures to significantly improve diagnostic accuracy and reliability in breast cancer detection. The custom-designed model offers a practical solution for early detection, potentially reducing false positives and false negatives, and improving patient outcomes through timely and accurate diagnosis.

PMID:39728907 | DOI:10.3390/tomography10120145

Tomography. 2024 Nov 30;10(12):1930-1946. doi: 10.3390/tomography10120140.

ABSTRACT

Background: Accurate reconstruction of internal temperature fields from surface temperature data is critical for applications such as non-invasive thermal imaging, particularly in scenarios involving small temperature gradients, like those in the human body. Methods: In this study, we employed 3D convolutional neural networks (CNNs) to predict internal temperature fields. The network's performance was evaluated under both ideal and non-ideal conditions, incorporating noise and background temperature variations. A physics-informed loss function embedding the heat equation was used in conjunction with statistical uncertainty during training to simulate realistic scenarios. Results: The CNN achieved high accuracy for small phantoms (e.g., 10 cm in diameter). However, under non-ideal conditions, the network's predictive capacity diminished in larger domains, particularly in regions distant from the surface. The introduction of physical constraints in the training processes improved the model's robustness in noisy environments, enabling accurate reconstruction of hot-spots in deeper regions where traditional CNNs struggled. Conclusions: Combining deep learning with physical constraints provides a robust framework for non-invasive thermal imaging and other applications requiring high-precision temperature field reconstruction, particularly under non-ideal conditions.

PMID:39728902 | DOI:10.3390/tomography10120140

Tomography. 2024 Nov 28;10(12):1915-1929. doi: 10.3390/tomography10120139.

ABSTRACT

Background: Assessment of skeletal maturity is a common clinical practice to investigate adolescent growth and endocrine disorders. The distal radius and ulna (DRU) maturity classification is a practical and easy-to-use scheme that was designed for adolescent idiopathic scoliosis clinical management and presents high sensitivity in predicting the growth peak and cessation among adolescents. However, time-consuming and error-prone manual assessment limits DRU in clinical application. Methods: In this study, we propose a multi-task learning framework with an attention mechanism for the joint segmentation and classification of the distal radius and ulna in hand X-ray images. The proposed framework consists of two sub-networks: an encoder-decoder structure with attention gates for segmentation and a slight convolutional network for classification. Results: With a transfer learning strategy, the proposed framework improved DRU segmentation and classification over the single task learning counterparts and previously reported methods, achieving an accuracy of 94.3% and 90.8% for radius and ulna maturity grading. Findings: Our automatic DRU assessment platform covers the whole process of growth acceleration and cessation during puberty. Upon incorporation into advanced scoliosis progression prognostic tools, clinical decision making will be potentially improved in the conservative and operative management of scoliosis patients.

PMID:39728901 | DOI:10.3390/tomography10120139