Nat Protoc. 2024 Dec 9. doi: 10.1038/s41596-024-01084-x. Online ahead of print.

ABSTRACT

Deep learning approaches have been increasingly applied to the discovery of novel chemical compounds. These predictive approaches can accurately model compounds and increase true discovery rates, but they are typically black box in nature and do not generate specific chemical insights. Explainable deep learning aims to 'open up' the black box by providing generalizable and human-understandable reasoning for model predictions. These explanations can augment molecular discovery by identifying structural classes of compounds with desired activity in lieu of lone compounds. Additionally, these explanations can guide hypothesis generation and make searching large chemical spaces more efficient. Here we present an explainable deep learning platform that enables vast chemical spaces to be mined and the chemical substructures underlying predicted activity to be identified. The platform relies on Chemprop, a software package implementing graph neural networks as a deep learning model architecture. In contrast to similar approaches, graph neural networks have been shown to be state of the art for molecular property prediction. Focusing on discovering structural classes of antibiotics, this protocol provides guidelines for experimental data generation, model implementation and model explainability and evaluation. This protocol does not require coding proficiency or specialized hardware, and it can be executed in as little as 1-2 weeks, starting from data generation and ending in the testing of model predictions. The platform can be broadly applied to discover structural classes of other small molecules, including anticancer, antiviral and senolytic drugs, as well as to discover structural classes of inorganic molecules with desired physical and chemical properties.

PMID:39653800 | DOI:10.1038/s41596-024-01084-x

Comput Biol Med. 2024 Dec 8;185:109541. doi: 10.1016/j.compbiomed.2024.109541. Online ahead of print.

ABSTRACT

Adenosine triphosphate plays a vital role in providing energy and enabling key cellular processes through interactions with binding proteins. The increasing amount of protein sequence data necessitates computational methods for identifying binding sites. However, experimental identification of adenosine triphosphate-binding residues remains challenging. To address the challenge, we developed a multi-window convolutional neural network architecture taking pre-trained protein language model embeddings as input features. In particular, multiple parallel convolutional layers scan for motifs localized to different window sizes. Max pooling extracts salient features concatenated across windows into a final multi-scale representation for residue-level classification. On benchmark datasets, our model achieves an area under the ROC curve of 0.95, significantly improving on prior sequence-based models and outperforming convolutional neural network baselines. This demonstrates the utility of pre-trained language models and multi-window convolutional neural networks for advanced sequence-based prediction of adenosine triphosphate-binding residues. Our approach provides a promising new direction for elucidating binding mechanisms and interactions from primary structure.

PMID:39653625 | DOI:10.1016/j.compbiomed.2024.109541

J Comput Chem. 2025 Jan 5;46(1):e27538. doi: 10.1002/jcc.27538.

ABSTRACT

Identifying interactions between drugs and targets is crucial for drug discovery and development. Nevertheless, the determination of drug-target binding affinities (DTAs) through traditional experimental methods is a time-consuming process. Conventional approaches to predicting drug-target interactions (DTIs) frequently prove inadequate due to an insufficient representation of drugs and targets, resulting in ineffective feature capture and questionable interpretability of results. To address these challenges, we introduce CGPDTA, a novel deep learning framework empowered by transfer learning, designed explicitly for the accurate prediction of DTAs. CGPDTA leverages the complementarity of drug-drug and protein-protein interaction knowledge through advanced drug and protein language models. It further enhances predictive capability and interpretability by incorporating molecular substructure graphs and protein pocket sequences to represent local features of drugs and targets effectively. Our findings demonstrate that CGPDTA not only outperforms existing methods in accuracy but also provides meaningful insights into the predictive process, marking a significant advancement in the field of drug discovery.

PMID:39653581 | DOI:10.1002/jcc.27538

JMIR AI. 2024 Dec 9;3:e55833. doi: 10.2196/55833.

ABSTRACT

Artificial intelligence (AI) has become commonplace in solving routine everyday tasks. Because of the exponential growth in medical imaging data volume and complexity, the workload on radiologists is steadily increasing. AI has been shown to improve efficiency in medical image generation, processing, and interpretation, and various such AI models have been developed across research laboratories worldwide. However, very few of these, if any, find their way into routine clinical use, a discrepancy that reflects the divide between AI research and successful AI translation. The goal of this paper is to give an overview of the intersection of AI and medical imaging landscapes. We also want to inform the readers about the importance of using standards in their radiology workflow and the challenges associated with deploying AI models in the clinical workflow. The main focus of this paper is to examine the existing condition of radiology workflow and identify the challenges hindering the implementation of AI in hospital settings. This report reflects extensive weekly discussions and practical problem-solving expertise accumulated over multiple years by industry experts, imaging informatics professionals, research scientists, and clinicians. To gain a deeper understanding of the requirements for deploying AI models, we introduce a taxonomy of AI use cases, supplemented by real-world instances of AI model integration within hospitals. We will also explain how the need for AI integration in radiology can be addressed using the Medical Open Network for AI (MONAI). MONAI is an open-source consortium for providing reproducible deep learning solutions and integration tools for radiology practice in hospitals.

PMID:39653370 | DOI:10.2196/55833

J Am Soc Echocardiogr. 2024 Dec 7:S0894-7317(24)00626-6. doi: 10.1016/j.echo.2024.11.010. Online ahead of print.

NO ABSTRACT

PMID:39653193 | DOI:10.1016/j.echo.2024.11.010

Lancet Oncol. 2024 Dec 6:S1470-2045(24)00599-0. doi: 10.1016/S1470-2045(24)00599-0. Online ahead of print.

ABSTRACT

BACKGROUND: Accurate detection of driver gene mutations is crucial for treatment planning and predicting prognosis for patients with lung cancer. Conventional genomic testing requires high-quality tissue samples and is time-consuming and resource-consuming, and as a result, is not available for most patients, especially those in low-resource settings. We aimed to develop an annotation-free Deep learning-enabled artificial intelligence method to predict GEne Mutations (DeepGEM) from routinely acquired histological slides.

METHODS: In this multicentre retrospective study, we collected data for patients with lung cancer who had a biopsy and multigene next-generation sequencing done at 16 hospitals in China (with no restrictions on age, sex, or histology type), to form a large multicentre dataset comprising paired pathological image and multiple gene mutation information. We also included patients from The Cancer Genome Atlas (TCGA) publicly available dataset. Our developed model is an instance-level and bag-level co-supervised multiple instance learning method with label disambiguation design. We trained and initially tested the DeepGEM model on the internal dataset (patients from the First Affiliated Hospital of Guangzhou Medical University, Guangzhou, China), and further evaluated it on the external dataset (patients from the remaining 15 centres) and the public TCGA dataset. Additionally, a dataset of patients from the same medical centre as the internal dataset, but without overlap, was used to evaluate the model's generalisation ability to biopsy samples from lymph node metastases. The primary objective was the performance of the DeepGEM model in predicting gene mutations (area under the curve [AUC] and accuracy) in the four prespecified groups (ie, the hold-out internal test set, multicentre external test set, TCGA set, and lymph node metastases set).

FINDINGS: Assessable pathological images and multigene testing information were available for 3697 patients who had biopsy and multigene next-generation sequencing done between Jan 1, 2018, and March 31, 2022, at the 16 centres. We excluded 60 patients with low-quality images. We included 3767 images from 3637 consecutive patients (1978 [54·4%] men, 1514 [41·6%] women, 145 [4·0%] unknown; median age 60 years [IQR 52-67]), with 1716 patients in the internal dataset, 1718 patients in the external dataset, and 203 patients in the lymph node metastases dataset. The DeepGEM model showed robust performance in the internal dataset: for excisional biopsy samples, AUC values for gene mutation prediction ranged from 0·90 (95% CI 0·77-1·00) to 0·97 (0·93-1·00) and accuracy values ranged from 0·91 (0·85-0·98) to 0·97 (0·93-1·00); for aspiration biopsy samples, AUC values ranged from 0·85 (0·80-0·91) to 0·95 (0·86-1·00) and accuracy values ranged from 0·79 (0·74-0·85) to 0·99 (0·98-1·00). In the multicentre external dataset, for excisional biopsy samples, AUC values ranged from 0·80 (95% CI 0·75-0·85) to 0·91 (0·88-1·00) and accuracy values ranged from 0·79 (0·76-0·82) to 0·95 (0·93-0·96); for aspiration biopsy samples, AUC values ranged from 0·76 (0·70-0·83) to 0·87 (0·80-0·94) and accuracy values ranged from 0·76 (0·74-0·79) to 0·97 (0·96-0·98). The model also showed strong performance on the TCGA dataset (473 patients; 535 slides; AUC values ranged from 0·82 [95% CI 0·71-0·93] to 0·96 [0·91-1·00], accuracy values ranged from 0·79 [0·70-0·88] to 0·95 [0·90-1·00]). The DeepGEM model, trained on primary region biopsy samples, could be generalised to biopsy samples from lymph node metastases, with AUC values of 0·91 (95% CI 0·88-0·94) for EGFR and 0·88 (0·82-0·93) for KRAS and accuracy values of 0·85 (0·80-0·88) for EGFR and 0·95 (0·92-0·96) for KRAS and showed potential for prognostic prediction of targeted therapy. The model generated spatial gene mutation maps, indicating gene mutation spatial distribution.

INTERPRETATION: We developed an AI-based method that can provide an accurate, timely, and economical prediction of gene mutation and mutation spatial distribution. The method showed substantial potential as an assistive tool for guiding the clinical treatment of patients with lung cancer.

FUNDING: National Natural Science Foundation of China, the Science and Technology Planning Project of Guangzhou, and the National Key Research and Development Program of China.

TRANSLATION: For the Chinese translation of the abstract see Supplementary Materials section.

PMID:39653054 | DOI:10.1016/S1470-2045(24)00599-0

Acta Psychol (Amst). 2024 Dec 8;251:104644. doi: 10.1016/j.actpsy.2024.104644. Online ahead of print.

ABSTRACT

Students' approaches to online technologies (SAOLT) have garnered significant support among scholars from a variety of fields, particularly in the field of science, technology, engineering, and mathematics education (STEM). The humanistic field needs to explore learners' approaches to target learning contexts, especially in the fields of computer-assisted language learning (CALL) and psycholinguistics. To fill this gap, the researchers explored 686 Iranian high school EFL learners' approaches to online language learning (OLL) with regard to their pedagogical, technical, and peer support, as well as the mediation role of online self-regulation. The partial least squares modeling approach's (PLS-SEM) reflective analysis validated the factorial structure of the conceptual study framework in online language learning and secondary education. The formative model found that instructional and peer support positively impacted language learners' deep and surface approaches to online language learning, thereby maximizing and minimizing their online presence and meaningful language learning. Furthermore, the mediation analysis revealed the significant moderating influence of learners' online self-regulation, which shifted the correlations from relating both learners' perceived support to their surface and deep learning approaches to only relating to the deep learning approach. Consequently, this study has pedagogical implications as it introduces a new conceptual framework to the CALL and psycholinguistic domains, specifically incorporating a new psychological factor related to language learners' complex dynamic systems, known as Language Learners' Approaches to Online Language Learning (OLLA). To do so, the researchers suggested that instructors should design more practical activities for theoretical subjects such as English language learning in online learning and shift online teaching from teacher-centered to peer-centered so as to foster their learners' deep learning approach to online language learning.

PMID:39652985 | DOI:10.1016/j.actpsy.2024.104644

JCO Clin Cancer Inform. 2024 Dec;8:e2400103. doi: 10.1200/CCI.24.00103. Epub 2024 Dec 9.

ABSTRACT

PURPOSE: Breast density is a widely established independent breast cancer risk factor. With the increasing utilization of digital breast tomosynthesis (DBT) in breast cancer screening, there is an opportunity to estimate volumetric breast density (VBD) routinely. However, current available methods extrapolate VBD from two-dimensional (2D) images acquired using DBT and/or depend on the existence of raw DBT data, which is rarely archived by clinical centers because of storage constraints.

METHODS: We retrospectively analyzed 1,080 nonactionable three-dimensional (3D) reconstructed DBT screening examinations acquired between 2011 and 2016. Reference tissue segmentations were generated using previously validated software that uses 3D reconstructed slices and raw 2D DBT data. We developed a deep learning (DL) model that segments dense and fatty breast tissue from background. We then applied this model to estimate %VBD and absolute dense volume (ADV) in cm3 in a separate case-control sample (180 cases and 654 controls). We created two conditional logistic regression models, relating each model-derived density measurement to likelihood of contralateral breast cancer diagnosis, adjusted for age, BMI, family history, and menopausal status.

RESULTS: The DL model achieved unweighted and weighted Dice scores of 0.88 (standard deviation [SD] = 0.08) and 0.76 (SD = 0.15), respectively, on the held-out test set, demonstrating good agreement between the model and 3D reference segmentations. There was a significant association between the odds of breast cancer diagnosis and model-derived VBD (odds ratio [OR], 1.41 [95 % CI, 1.13 to 1.77]; P = .002), with an AUC of 0.65 (95% CI, 0.60 to 0.69). ADV was also significantly associated with breast cancer diagnosis (OR, 1.45 [95% CI, 1.22 to 1.73]; P < .001) with an AUC of 0.67 (95% CI, 0.62 to 0.71).

CONCLUSION: DL-derived density measures derived from 3D reconstructed DBT images are associated with breast cancer diagnosis.

PMID:39652797 | DOI:10.1200/CCI.24.00103

PLoS One. 2024 Dec 9;19(12):e0313329. doi: 10.1371/journal.pone.0313329. eCollection 2024.

ABSTRACT

Cone snails are venomous marine gastropods comprising more than 950 species widely distributed across different habitats. Their conical shells are remarkably similar to those of other invertebrates in terms of color, pattern, and size. For these reasons, assigning taxonomic signatures to cone snail shells is a challenging task. In this report, we propose an ensemble learning strategy based on the combination of Random Forest (RF) and XGBoost (XGB) methods. We used 47,600 cone shell images of uniform size (224 x 224 pixels), which were split into an 80:20 train-test ratio. Prior to performing subsequent operations, these images were subjected to pre-processing and transformation. After applying a deep learning approach (Visual Geometry Group with a 16-layer deep model architecture) for feature extraction, model specificity was further assessed by including multiple related and unrelated seashell images. Both classifiers demonstrated comparable recognition ability on random test samples. The evaluation results suggested that RF outperformed XGB due to its high accuracy in recognizing Conus species, with an average precision of 95.78%. The area under the receiver operating characteristic curve was 0.99, indicating the model's optimal performance. The learning and validation curves also demonstrated a robust fit, with the training score reaching 1 and the validation score gradually increasing to 95 as more data was provided. These values indicate a well-trained model that generalizes effectively to validation data without significant overfitting. The gradual improvement in the validation score curve is crucial for ensuring model reliability and minimizing the risk of overfitting. Our findings revealed an interactive visualization. The performance of our proposed model suggests its potential for use with datasets of other mollusks, and optimal results may be achieved for their categorization and taxonomical characterization.

PMID:39652613 | DOI:10.1371/journal.pone.0313329

J Christ Nurs. 2025 Jan-Mar 01;42(1):22-27. doi: 10.1097/CNJ.0000000000001226. Epub 2024 Dec 9.

ABSTRACT

The integration of faith and learning (IFL) in nursing has deep historical roots. However, barriers have developed to successful IFL in Christian higher education. A biblically based concept curriculum (BBCC) is proposed that emphasizes deep learning, critical thinking, and student-centered learning. Examples are provided of BBCC curricular integration, pedagogical methods, and assessments related to the IFL. Results of student evaluations before and after implementation of the BBCC identified an increase in students' perceptions of IFL in nursing courses.

PMID:39652482 | DOI:10.1097/CNJ.0000000000001226

Oral Radiol. 2024 Dec 9. doi: 10.1007/s11282-024-00794-y. Online ahead of print.

ABSTRACT

OBJECTIVES: Deep learning has revolutionized image analysis for dentistry. Automated segmentation of dental radiographs is of great importance towards digital dentistry. The performance of deep learning models heavily relies on the quality and diversity of the training data. Data augmentation is a widely used technique implemented in machine learning and deep learning to artificially increase the size and diversity of a training dataset by applying various transformations to the original data.

METHODS: This work aims to automatically segment implants, prostheses, and fillings in panoramic images using 9 different deep learning segmentation models. Later, it explores the effect of data augmentation methods on segmentation performance of the models. Eight different data augmentation techniques are examined. Performance is evaluated by well-accepted metrics such as intersection over union (IoU) and Dice coefficient.

RESULTS: While averaging the segmentation results for the three classes, IoU varies between 0.62 and 0.82 while Dice score is between 0.75 and 0.9 among deep learning models used. Augmentation techniques provided performance improvements of up to 3.37%, 5.75% and 8.75% for implant, prosthesis and filling classes, respectively.

CONCLUSIONS: Findings reveal that choosing optimal augmentation strategies depends on both model architecture and dental structure type.

PMID:39652261 | DOI:10.1007/s11282-024-00794-y

Int J Comput Assist Radiol Surg. 2024 Dec 9. doi: 10.1007/s11548-024-03301-0. Online ahead of print.

ABSTRACT

PURPOSE: Awake craniotomy presents a unique opportunity to map and preserve critical brain functions, particularly speech, during tumor resection. The ability to accurately assess linguistic functions in real-time not only enhances surgical precision, but also contributes significantly to improving postoperative outcomes. However, today, its evaluation is subjective as it relies on a clinician's observations only. This paper explores the use of a deep learning based model for the objective assessment of speech arrest and speech impairments during awake craniotomy.

METHODS: We extracted 1883 3-second audio clips containing the patient's response following direct electrical stimulation from 23 awake craniotomies recorded from two operating rooms of the Tokyo Women's Medical University Hospital (Japan) and two awake craniotomies recorded from the University Hospital of Brest (France). A Wav2Vec2-based model has been trained and used to detect speech arrests and speech impairments. Experiments were performed with different datasets settings and preprocessing techniques and the performances of the model were evaluated using the F1-score.

RESULTS: The F1-score was 84.12% when the model was trained and tested on Japanese data only. In a cross-language situation, the F1-score was 74.68% when the model was trained on Japanese data and tested on French data.

CONCLUSIONS: The results are encouraging even in a cross-language situation but further evaluation is required. The integration of preprocessing techniques, in particular noise reduction, improved the results significantly.

PMID:39652158 | DOI:10.1007/s11548-024-03301-0

Med Phys. 2024 Dec 9. doi: 10.1002/mp.17545. Online ahead of print.

ABSTRACT

BACKGROUND: Ultrasound (US) has great potential for application in computer-assisted orthopedic surgery (CAOS) due to its non-radiative, cost-effective, and portable traits. However, bone segmentation from low-quality US images has been challenging. Traditional segmentation methods cannot achieve satisfactory results due to their high customization and dependence on bone morphology. Existing deep learning-based methods make it difficult to ensure efficient and accurate segmentation due to the ignorance of prior knowledge of bone features during feature learning.

PURPOSE: This paper aims to systematically investigate feature extraction and segmentation methodologies of bone US images and then proposes an innovative convolutional neural network to address the need for precise and efficient bone structure extraction in CAOS.

METHODS: This paper has proposed a dual-decoder banded convolutional attention network (BCA-Net), which takes the raw US image as input and simplified U-Net as the baseline network. Multiscale banded convolution kernels are employed internally in the BCA-Net model, leveraging the prior knowledge that bone surfaces in US images are exhibited as bright bands of a few millimeters in width. Additionally, a shared encoder to extract input features and two independent decoders to generate outputs for the bone surface and bone shadow mask are integrated into the BCA-Net model, leveraging the prior knowledge that US bone surfaces manifest low-intensity hollow shadows below. Then, a new task consistency loss is introduced that can utilize inter-task dependency fully and enhance the performance of our model. In the network construction process, a dataset containing 1623 sets of US images was adopted, and a five-fold cross-validation strategy was divided into the training and validation sets for the model's training and validation. Many vital metrics were introduced to comprehensively evaluate the model performance, including overlap, edge distance, area under curve, and model efficiency. Finally, the model performance was subjected to a confidence interval, Tukey's honest significant difference, and Cohen's d statistics at a significance level (5%) to ensure the accuracy and reliability of the obtained findings.

RESULTS: The experimental results show that the BCA-Net model performs well in the bone surface segmentation task. Its average Dice coefficient reaches 87.51%, 4.04% higher than U-Net's, proving its superior bone surface segmentation accuracy. Meanwhile, the average distance error is 0.2 mm, 0.33 mm lower than U-Net's, highlighting its accuracy in detail capture and boundary recognition. Using a confidence distance threshold of 1.02 mm, the Dice coefficient of the BCA-Net model exceeds 98%, an improvement of 1.87% over U-Net's, which is highly consistent with manual labeling. The BCA-Net model achieves a statistical significance of p-values < 0.05 in the above accuracy comparisons. In addition, the BCA-Net model has a small parameter count (5.58 M) and high computational efficiency (35.85 frames per second), further validating its excellent potential in bone surface segmentation tasks.

CONCLUSIONS: The proposed method achieves excellent performance with high accuracy and efficiency, aligning well with clinical requirements and holding excellent potential for advancing the utilization of US images in CAOS.

PMID:39651711 | DOI:10.1002/mp.17545

J Microsc. 2024 Dec 9. doi: 10.1111/jmi.13376. Online ahead of print.

ABSTRACT

Conventional optical microscopy imaging of obligate intracellular bacteria is hampered by the small size of bacterial cells, tight clustering exhibited by some bacterial species and challenges relating to labelling such as background from host cells, a lack of validated reagents, and a lack of tools for genetic manipulation. In this study, we imaged intracellular bacteria from the species Orientia tsutsugamushi (Ot) using five different fluorescence microscopy techniques: standard confocal, Airyscan confocal, instant Structured Illumination Microscopy (iSIM), three-dimensional Structured Illumination Microscopy (3D-SIM) and Stimulated Emission Depletion Microscopy (STED). We compared the ability of each to resolve bacterial cells in intracellular clumps in the lateral (xy) axis, using full width half-maximum (FWHM) measurements of a labelled outer membrane protein (ScaA) and the ability to detect small, outer membrane vesicles external to the cells. Comparing the techniques readily available to us (above), 3D-SIM microscopy, in combination with the shortest-wavelength dyes, was found overall to give the best lateral resolution. We next compared the ability of each technique to sufficiently resolve bacteria in the axial (z) direction and found 3D-STED to be the most successful method for this. We then combined this 3D-STED approach with a custom 3D cell segmentation and analysis pipeline using the open-source, deep learning software, Cellpose to segment the cells and subsequently the commercial software Imaris to analyse their 3D shape and size. Using this combination, we demonstrated differences in bacterial shape, but not their size, when grown in different mammalian cell lines. Overall, we compare the advantages and disadvantages of different super-resolution microscopy techniques for imaging this cytoplasmic obligate intracellular bacterium based on the specific research question being addressed.

PMID:39651611 | DOI:10.1111/jmi.13376

Big Data. 2024 Dec 9. doi: 10.1089/big.2024.0071. Online ahead of print.

ABSTRACT

Liver cirrhosis stands as a prominent contributor to mortality, impacting millions across the United States. Enabling health care providers to predict early mortality among patients with cirrhosis holds the potential to enhance treatment efficacy significantly. Our hypothesis centers on the correlation between mortality and laboratory test results along with relevant diagnoses in this patient cohort. Additionally, we posit that a deep learning model could surpass the predictive capabilities of the existing Model for End-Stage Liver Disease score. This research seeks to advance prognostic accuracy and refine approaches to address the critical challenges posed by cirrhosis-related mortality. This study evaluates the performance of an artificial neural network model for liver disease classification using various training dataset sizes. Through meticulous experimentation, three distinct training proportions were analyzed: 70%, 80%, and 90%. The model's efficacy was assessed using precision, recall, F1-score, accuracy, and support metrics, alongside receiver operating characteristic (ROC) and precision-recall (PR) curves. The ROC curves were quantified using the area under the curve (AUC) metric. Results indicated that the model's performance improved with an increased size of the training dataset. Specifically, the 80% training data model achieved the highest AUC, suggesting superior classification ability over the models trained with 70% and 90% data. PR analysis revealed a steep trade-off between precision and recall across all datasets, with 80% training data again demonstrating a slightly better balance. This is indicative of the challenges faced in achieving high precision with a concurrently high recall, a common issue in imbalanced datasets such as those found in medical diagnostics.

PMID:39651607 | DOI:10.1089/big.2024.0071

Curr Eye Res. 2024 Dec 9:1-6. doi: 10.1080/02713683.2024.2430212. Online ahead of print.

ABSTRACT

PURPOSE: To examine the performance of deep-learning models that predicts the visual acuity after cataract surgery using preoperative clinical information and color fundus photography (CFP).

METHODS: We retrospectively collected the age, sex, and logMAR preoperative best corrected visual acuity (preoperative-BCVA) and CFP from patients who underwent cataract surgeries from 2020 to 2021 at National Taiwan University Hospital. Feature extraction of CFP was performed using a pre-existing image classification model, Xception. The CFP-extracted features and pre-operative clinical information were then fed to a downstream neural network for final prediction. We assessed the model performance by calculating the mean absolute error (MAE) between the predicted and the true logMAR of postoperative BCVA. A nested 10-fold cross-validation was performed for model validation.

RESULTS: A total of 673 fundus images from 446 patients were collected. The mean preoperative BCVA and postoperative BCVA was 0.60 ± 0.39 and 0.14 ± 0.18, respectively. The model using age and sex as predictors achieved an MAE of 0.121 ± 0.016 in postoperative BCVA prediction. The inclusion of CFP as additional predictor in the model (predictors: age, sex and CFP) did not further improve the predictive performance (MAE = 0.120 ± 0.023, p = 0.375), while adding the preoperative BCVA as an additional predictor resulted in a 4.13% improvement (predictors: age, sex and preoperative BCVA, MAE = 0.116 ± 0.016, p = 0.013).

CONCLUSIONS: Our multimodal models including both CFP and clinical information achieved excellent accuracy in predicting BCVA after cataract surgery, while the learning models input with only clinical information performed similarly. Future studies are needed to clarify the effects of multimodal input on this task.

PMID:39651583 | DOI:10.1080/02713683.2024.2430212

Int J Comput Dent. 2024 Dec 9;27(4):401-411. doi: 10.3290/j.ijcd.b5870240.

ABSTRACT

AIM: The objective of the present study was to investigate the clinical understanding and reasoning abilities of large language models (LLMs); namely, ChatGPT, GPT-4, and New Bing, by evaluating their performance in the NDLE (National Dental Licensing Examination) in China.

MATERIALS AND METHODS: Questions from the NDLE from 2020 to 2022 were selected based on subject weightings. Standardized prompts were utilized to regulate the output of LLMs for acquiring more precise answers. The performance of each model across each subject category and for the subjects overall was analyzed employing the McNemar's test.

RESULTS: The percentage scores obtained by ChatGPT, GPT-4, and New Bing were 42.6% (138/324), 63.0% (204/324), and 72.5% (235/324), respectively. Significant variance was seen between the performance of New Bing compared with ChatGPT and GPT-4. GPT-4 and New Bing outperformed ChatGPT across all subjects, with New Bing surpassing GPT-4 in most subjects.

CONCLUSION: GPT-4 and New Bing exhibited promising capabilities in the NDLE. However, their performance in specific subjects such as prosthodontics and oral and maxillofacial surgery requires improvement. This performance gap can be attributed to limited dental training data and the inherent complexity of these subjects.

PMID:39651568 | DOI:10.3290/j.ijcd.b5870240

Interface Focus. 2024 Dec 6;14(6):20240024. doi: 10.1098/rsfs.2024.0024. eCollection 2024 Dec 6.

ABSTRACT

Functional magnetic resonance imaging (fMRI) captures rich physiological and neuronal information, offering insight into neurofluid dynamics, vascular health and waste clearance. Accurate cerebral vessel segmentation could greatly facilitate fluid dynamics research in fMRI. However, existing vessel identification methods, such as magnetic resonance angiography or deep-learning-based segmentation on structural MRI, cannot reliably locate cerebral vessels in fMRI space due to misregistration from inherent fMRI distortions. To address this challenge, we developed a data-driven, automatic segmentation of cerebral vessels directly within fMRI space. This approach identified large cerebral arteries and the superior sagittal sinus (SSS) by leveraging these vessels' distinct pulsatile signal patterns during the cardiac cycle. The method was validated in a local dataset by comparing it to ground truth cerebral artery and SSS segmentations. Using the Human Connectome Project (HCP) ageing dataset, the method's reproducibility was tested on 422 participants aged 36-90, each with four repeated fMRI scans. The method demonstrated high reproducibility, with an intraclass correlation coefficient > 0.7 in both cerebral artery and SSS segmentation volumes. This study demonstrates that large cerebral arteries and SSS can be reproducibly and automatically segmented in fMRI datasets, facilitating reliable fluid dynamics investigation in these regions.

PMID:39649451 | PMC:PMC11620823 | DOI:10.1098/rsfs.2024.0024

Npj Imaging. 2024;2(1):51. doi: 10.1038/s44303-024-00055-x. Epub 2024 Dec 4.

ABSTRACT

Voltage imaging is a powerful technique for studying neuronal activity, but its effectiveness is often constrained by low signal-to-noise ratios (SNR). Traditional denoising methods, such as matrix factorization, impose rigid assumptions about noise and signal structures, while existing deep learning approaches fail to fully capture the rapid dynamics and complex dependencies inherent in voltage imaging data. Here, we introduce CellMincer, a novel self-supervised deep learning method specifically developed for denoising voltage imaging datasets. CellMincer operates by masking and predicting sparse pixel sets across short temporal windows and conditions the denoiser on precomputed spatiotemporal auto-correlations to effectively model long-range dependencies without large temporal contexts. We developed and utilized a physics-based simulation framework to generate realistic synthetic datasets, enabling rigorous hyperparameter optimization and ablation studies. This approach highlighted the critical role of conditioning on spatiotemporal auto-correlations, resulting in an additional 3-fold SNR gain. Comprehensive benchmarking on both simulated and real datasets, including those validated with patch-clamp electrophysiology (EP), demonstrates CellMincer's state-of-the-art performance, with substantial noise reduction across the frequency spectrum, enhanced subthreshold event detection, and high-fidelity recovery of EP signals. CellMincer consistently outperforms existing methods in SNR gain (0.5-2.9 dB) and reduces SNR variability by 17-55%. Incorporating CellMincer into standard workflows significantly improves neuronal segmentation, peak detection, and functional phenotype identification, consistently surpassing current methods in both SNR gain and consistency.

PMID:39649342 | PMC:PMC11618097 | DOI:10.1038/s44303-024-00055-x

Phys Imaging Radiat Oncol. 2024 Nov 16;32:100676. doi: 10.1016/j.phro.2024.100676. eCollection 2024 Oct.

ABSTRACT

BACKGROUND AND PURPOSE: This study aimed at investigating the dosimetric impact on organs at risk, when the left-sided internal mammary lymph nodes (IMN) were delineated with two interpretations of the same guideline.

MATERIALS AND METHODS: The cohort consisted of 95 left-sided breast cancer patients with indication for irradiation of the CTVn_IMN treated at the Netherlands Cancer Institute (NKI). The NKI interpretation of the ESTRO guidelines was in the clinical structure sets (CTVn_IMN_NKI). A deep learning model was used as second interpretation of the guideline, based on a Danish consensus interpretation (CTVn_IMN_DK). The geometrical similarity was evaluated with the Dice Similarity Coefficient (DSC), volume, width, distance to sternal bone (SB) and maximum distance between the interpretations in the medial direction. Treatment plans were generated for both CTVn_IMNs. Mean heart dose (MHD) was correlated with the geometrical metrics.

RESULTS: 62 patients were eligible for analysis. The geometric comparison showed a median volume of 9.59 ml/7.19 ml for the CTVN_IMN_NKI/CTVn_IMN_DK along with a median DSC of 0.63. The width and distance from SB were significantly different, with a median width of 18.2 mm/14.7 mm and distance to SB of 3.4 mm/5.1 mm for CTVn_IMN_NKI/CTVn_IMN_DK. The MHD was significantly higher with the CTVn_IMN_NKI. The strongest correlation was found between MHD and maximum medial difference between the CTVn_IMN in slices where the heart was present.

CONCLUSIONS: Differences in interpretations of the CTVn_IMN delineation guidelines were found, resulting in significant differences in MHD. For the individual patients, the dosimetric differences may impact treatment decisions, underscoring the need for strong consensus across borders.

PMID:39649154 | PMC:PMC11625340 | DOI:10.1016/j.phro.2024.100676