Test Dataset Research Articles

Mapping fine-scale socioeconomic inequality using machine learning and remotely sensed data

Abstract Limited and missing socioeconomic data have made it nearly impossible to measure or estimate inequality consistently at fine spatiotemporal and jurisdictional scales, especially for lower- and middle-income countries (L&MICs). We deploy a novel data harmonization method that combines existing household survey data with freely available remotely sensed data and machine learning techniques to generate fine-scale socioeconomic inequality measures across spatial and temporal scales for India. Our manuscript makes three important contributions. One, it identifies key remote sensing datasets that, in combination with nighttime luminosity, improve its predictive power to estimate measures of socioeconomic inequality; Two, it offers an analytical approach that reliably estimates the uneven distribution of socioeconomic conditions by harmonizing household assets and socio-demographic information that remotely sensed data at the village or similar geographic levels represent – the results achieve > 84% prediction accuracy. Finally, it leverages a spatially cross-validated machine learning model with training and test datasets from two successive Demographic & Health Surveys (DHS) to demonstrate how data gaps in socioeconomic inequality at subnational levels can be addressed. Our replicable approach has the potential to improve global inequality data, thereby supporting research and applications aiming to reduce socioeconomic inequality in the context of the Sustainable Development Goals (SDGs).

Development and validation of a novel bleeding risk prediction tool for aspirin users with a low body mass index

Aspirin is commonly utilized in the management and prevention of various diseases. However, in specific individuals, particularly those with low body mass index (BMI), aspirin can elevate the risk of bleeding. Achieving a delicate equilibrium between the desirable antiplatelet effects and potential bleeding complications is a notable consideration. The objective of this study was to create a novel bleeding risk prediction tool for aspirin users with a low BMI. A total of 2436 aspirin users with a low BMI were included in this study conducted at the Affiliated Dongyang Hospital of Wenzhou Medical University. Patient data, comprising demographics, clinical characteristics, comorbidities, medical history, and laboratory tests, were collected. The patients were randomly divided into two groups, with a 7:3 ratio, for model development and internal validation purposes. The identification of clinically significant features associated with bleeding was achieved through the utilization of the Least Absolute Shrinkage and Selection Operator (LASSO) regression and boruta analysis. Subsequently, these important features underwent multivariate logistic regression analysis. Based on independent bleeding risk factors, a logistic regression model was constructed and presented as a nomogram. Model performance was evaluated using metrics such as the area under the curve (AUC), calibration curves, decision curve analysis (DCA), clinical impact curve (CIC), and net reduction curve (NRC) in both the training and testing sets. LASSO analysis identified two clinical features, while Boruta analysis identified nine clinical features out of a total of 21 features. Subsequent multivariate logistic regression analysis selected significant independent risk factors. The boruta model, which demonstrated the highest AUC, consisted of six clinical variables: hemoglobin, platelet count, previous bleeding, tumor, smoke, and diabetes mellitus. These variables were integrated into a visually represented nomogram. The model exhibited an AUC of 0.832 (95% CI: 0.788–0.875) in the training dataset and 0.775 (95% CI: 0.698–0.853) in the test dataset, indicating excellent discriminatory performance. Calibration curve analysis revealed close alignment with the ideal curve. Furthermore, DCA, CIC, and NRC demonstrated favorable clinical net benefit for the model. This study has successfully created a novel risk prediction tool specifically designed for aspirin users with a low BMI. This tool enables the stratification of low BMI patients based on their anticipated bleeding risk.

Predicting cardiovascular outcomes in Chinese patients with type 2 diabetes by combining risk factor trajectories and machine learning algorithm: a cohort study

BackgroundCardiovascular complications are major concerns for Chinese patients with type 2 diabetes. Accurately predicting these risks remains challenging due to limitations in traditional risk models. We aimed to develop a dynamic prediction model using machine learning and longitudinal trajectories of cardiovascular risk factors to improve prediction accuracy.MethodsWe included 16,378 patients from the Kailuan cohort, splitting them into training and testing datasets. Using baseline characteristics and changes over a four-year observation period, we developed the ML-CVD-C (Machine Learning Cardiovascular Disease in Chinese) score to predict 10-year cardiovascular risk, including cardiovascular death, nonfatal myocardial infarction, and stroke. We compared the discrimination and calibration of ML-CVD-C with models using only baseline variables (ML-CVD-C [base]), China-PAR (Prediction for ASCVD Risk in China), and PREVENT (Predict Risk of cardiovascular disease EVENTs). Risk stratification improvements were assessed through net reclassification improvement (NRI) and integrated discrimination improvement (IDI). Transition analysis examined the changes in risk stratification over time.ResultsThe ML-CVD-C score achieved a C-index of 0.80 (95% CI: 0.78–0.82) in the testing cohort, significantly outperforming the ML-CVD-C (base) score, China-PAR, and PREVENT, which had C-index values of 0.62–0.65. ML-CVD-C also provided more accurate cardiovascular risk estimates, though all models tended to overestimate the prevalence of high-risk cases. Stratification by the ML-CVD-C score showed substantial improvement, with NRI gains of 57.7%, 44.1%, and 47.3%, and IDI gains of 10.1%, 7.9%, and 8.4% compared to the other three scores. Both the trajectory and machine learning algorithm contributed significantly to the enhancement of model performance. Transition analysis revealed that participants who remained in the same risk category or were reclassified to a lower category exhibited 22% and 86% reductions in cardiovascular risk compared to those reclassified to a higher risk category during the observation period.ConclusionsThe ML-CVD-C model, incorporating dynamic cardiovascular risk trajectories and a machine learning algorithm, significantly improves risk prediction accuracy for Chinese patients with diabetes. This model may serve as a valuable tool for more personalized cardiovascular risk management in type 2 diabetes.

A ConvLSTM-based model for predicting thermal damage during laser interstitial thermal therapy.


Accurate prediction of thermal damage extent is essential for effective and precise thermal therapy, especially in brain laser interstitial thermal therapy (LITT). Immediate postoperative contrast-enhanced T1-weighted imaging (CE-T1WI) is the primary method for clinically assessing in vivo thermal damage after image-guided LITT. CE-T1WI reveals a hyperintense enhancing rim surrounding the target lesion, which serves as a key radiological marker for evaluating the thermal damage extent. Although widely used in clinical practice, traditional thermal damage models rely on empirical parameters from in vitro experiments, which can lead to inaccurate predictions of thermal damage in vivo. Additionally, these models predict only two tissue states (damaged or undamaged), failing to capture three tissue states observed on post-CE-T1WI images, highlighting the need for improved thermal damage prediction methods. This study proposes a novel Convolutional Long Short-Term Memory (ConvLSTM)-based model that utilizes intraoperative temperature distribution history data measured by magnetic resonance temperature imaging (MRTI) during LITT to predict the enhancing rim on post-CE-T1WI images. This method was implemented and evaluated on retrospective data from 56 patients underwent brain LITT.
Main results:
The proposed model effectively predicts the enhancing rim on postoperative images, achieving an average Dice Similarity Coefficient (DSC) of 0.82 (±0.063) on the test dataset. Furthermore, it generates real-time predicted thermal damage area variation trends that closely resemble those of the traditional thermal damage model, suggesting potential for real-time prediction of thermal damage extent. This method could provide a valuable tool for visualizing and assessing intraoperative thermal damage extent.&#xD.

Intelligent leaching of Zn and Mn from spent disposable batteries to avoid traditional optimizing experiments.

Spent disposable Zn-Mn and Zn-C batteries are important resources for recycling. Acid leaching is the crucial step in the hydrometallurgy process for recycling Zn and Mn from these spent Zn-based batteries. However, to obtain the optimal leaching efficiency, the uncontrollable components in waste feed and various leaching parameters cause numerous replicated optimal experiments, increasing the recovery cost and environmental risks. To solve the issues, we employed machine learning (ML) techniques to construct models to predict Zn and Mn leaching from spent disposable batteries without optimizing experiments. Among four ML algorithms tested, the extreme gradient boosting demonstrated superior predictive performance, achieving an R2 of 0.85-0.98 across the training, test, and verification datasets. An analysis of feature importance indicated that the particle size, waste composition, acid concentration, temperature, and time affected the metal leaching most. This study also revealed the interaction effects of the waste properties and leaching process on the metal leaching. Furthermore, we created a user-friendly graphical user interface (GUI) that enables quick acquisition of metal leaching results, requiring only the measurement of waste particle size and component. Finally, experimental verification confirmed the practicability of the GUI. This study achieves intelligent metal leaching from spent batteries and overcomes the high recovery cost and environmental risks associated with traditional experimental optimizing methods.

Leveraging artificial intelligence to promote COVID-19 appropriate behaviour in a healthcare institution from north India: A feasibility study

Background & Objectives Non-pharmacological interventions (NPI) were crucial in curbing the initial COVID-19 pandemic waves, but compliance was difficult. The primary aim of this study was to assess the changes in compliance with NPIs in healthcare settings using Artificial intelligence (AI) and examine the barriers and facilitators of using AI systems in healthcare. Methods A pre-post-intervention study was conducted in a north-Indian hospital between April and July 2022. YOLO-V5 and 3D Cartesian distance algorithm-based AI modules were used to ascertain compliance through several parameters like confidence threshold, intersection-over-union threshold, image size, distance threshold (6 feet), and 3D Euclidean Distance estimation. Validation was done by evaluating model performance on a labelled test dataset, and accuracy was 91.3 per cent. Interventions included daily sensitization and health education for the hospital staff and visitors, display of information, education and communication (IEC) materials, and administrative surveillance. In-depth interviews were conducted with the stakeholders to assess the feasibility issues. Flagged events during the three phases were compared using One-way ANOVA tests in SPSS. Results Higher social distancing (SD) compliance events were flagged by the module in the intervention phase compared to the pre-intervention and post-intervention phases (P<0.05). Mask non-compliance was significantly lower (P <0.05) in the pre-intervention phase and highest in the post-intervention phase, with varied differences between different intervention phases in the registration hall and medicine out-patient department (OPD). The modules’ data safety, transfer, and cost were the most common concerns. Interpretation & conclusions AI can supplement our efforts against the pandemic and offer indispensable help with minimal feasibility issues that can be resolved through adequate sensitization and training.

Interpretable Radiomics Model Predicts Nanomedicine Tumor Accumulation Using Routine Medical Imaging.

Accurately predicting nanomedicine accumulation is critical for guiding patient stratification and optimizing treatment strategies in the context of precision medicine. However, non-invasive prediction of nanomedicine accumulation remains challenging, primarily due to the complexity of identifying relevant imaging features that predict accumulation. Here, a novel non-invasive method is proposed that utilizes standard-of-care medical imaging modalities, including computed tomography and ultrasound, combined with a radiomics-based model to predict nanomedicine accumulation in tumor. The model is validated using a test dataset consisting of seven tumor xenografts in mice and three sizes of gold nanoparticles, achieving an area under the receiver operating characteristic curve of 0.851. The median accumulation levels of tumors predicted as "high accumulators" are 2.69 times greater than those predicted as "low accumulators". Analysis of this machine-learning-driven interpretable radiomics model revealed imaging features that are strongly correlated with dense stroma, a recognized biological barrier to effective nanomedicine delivery. Radiomics-based prediction of tumor accumulation holds promise for stratifying patient and enabling precise tailoring of nanomedicine treatment strategies.

Genome data based deep learning identified new genes predicting pharmacological treatment response of attention deficit hyperactivity disorder

Although the efficacy of pharmacy in the treatment of attention deficit/hyperactivity disorder (ADHD) has been well established, the lack of predictors of treatment response poses great challenges for personalized treatment. The current study employed a comprehensive approach, combining genome-wide association analyses (GWAS) and deep learning (DL) methods, to elucidate the genetic underpinnings of pharmacological treatment response in ADHD. Based on genotype data of medication-naïve patients with ADHD who received pharmacological treatments for 12 weeks, the current study performed GWAS using the percentage changes in ADHD-RS score as phenotype. Then, DL models were constructed to predict percentage changes in symptom scores using genetic variants selected based on four different genome-wide P thresholds (E-02, E-03, E-04, E-05) as inputs. The current GWAS results identified two significant loci (rs10880574, P = 2.39E-09; rs2000900, P = 3.31E-09) which implicated two genes, TMEM117 and MYO5B, that were primarily associated with both brain- and gut-related disorders. The convolutional neural network (CNN) model, using variants with genome-wide P values less than E-02 (5516 SNPs), demonstrated the best performance with mean squared error (MSE) equals 0.012 (Accuracy = 0.83; Sensitivity = 0.90; Specificity = 0.75) in the validation dataset, 0.081 in an independent test dataset (Acc = 0.61, Sensitivity = 0.81; Specificity = 0.26). Notably, the variant that contributed most to the CNN model was NKAIN2, an ADHD-related gene, which is also associated with metabolic processes. To conclude, the integration of GWAS and DL methods revealed new genes contribute to ADHD pharmacological treatment responses, and underscored the interplay between neural systems and metabolic processes, potentially providing critical insights into precision treatment. Furthermore, our CNN model exhibited good performance in an independent dataset, encouraged future studies and implied potential clinical applications.

Innovating bio-lubricant design: utilizing Mordred descriptors for model development and thermal stability prediction of novel compounds

Purpose Developing bio-lubricants for high-temperature applications, such as engine operations, requires reliable thermal stability assessment. This study aims to create a predictive tool to evaluate thermal stability using existing organic compounds data. The model predicts the onset temperature (T onset) of bio-lubricant candidates derived from epoxidation with various amine nucleophiles, enhancing initial selection efficiency and reducing reliance on traditional trial-and-error methods. Design/methodology/approach Using the Python library Mordred, molecular descriptors were calculated from the SMILES structures of 126 compounds with known T onsets. These descriptors were inputs for machine learning models predicting the thermal stability of bio-lubricants. The models were evaluated on training and test data sets and validated with novel synthesized compounds. Findings The predictive model showed strong performance on the test data set, with an R² of 0.93 and a root mean square error of 17.77. In external validation, the model estimated the thermal stability of a novel bio-lubricant compound (MA-EPO) at 290.7°C, closely matching the actual T onset of approximately 300°C. Originality/value This research introduces a method for predicting the thermal stability of bio-lubricants using machine learning and open-source libraries. This approach significantly advances the field by improving the efficiency of bio-lubricant development for high-temperature applications. It provides a cost-effective and time-saving alternative to traditional experimental methods, serving as a valuable resource for researchers and developers in the bio-lubricant industry.

Comparison of algorithms for the recognition of ChatGPT paraphrased texts

The rapid development of artificial intelligence, especially AI assistants, is leading to new forms of plagiarism that are difficult to detect using existing methods. Paraphrasing tools make this problem even more complex and challenging especially in minor languages with inadequate resources and tools. This study explores strategies to help detect plagiarism generated by ChatGPT 4.0 and altered by paraphrasing tools. We propose two new datasets consisting of abstracts of doctoral theses in English and Serbian. Both datasets were subjected to ChatGPT paraphrasing, which allowed us to form two classes of texts: human-written and AI-generated, i.e., AI-paraphrased. We then comprehensively compare 19 widely used classification algorithms based on two feature sets: word unigrams and character multigrams. In addition, we compare these to the results of a commercially available pre-trained ChatGPT content detector, ZeroGPT. The results on the English corpus turn out to be very accurate, achieving an accuracy of 95% or more. In contrast, the results on the Serbian corpus were less accurate, achieving an accuracy of just over 85%. Syntax analysis of the training datasets has shown that in Serbian GPT-paraphrased texts, 33.2% of sentences remain the same, and they are found in 63% of documents. GPT-paraphrased English texts showed that 3.2% of sentences remain the same, and they are found in 16% of documents. Syntax analysis of the test datasets has shown that the change of the model temperature influences syntactic features (average number of words and sentences) in English texts and slightly or not in Serbian texts. We attribute all these differences to GPT’s lower paraphrasing ability in minor languages such as Serbian. Presented findings underscore the necessity for making persistent effort in developing tools made for detecting AI-paraphrased texts in academic and professional settings, particularly for minor languages with limited NLP resources, to preserve content integrity and authenticity.

Enhancing deep learning methods for brain metastasis detection through cross-technique annotations on SPACE MRI

BackgroundGadolinium-enhanced “sampling perfection with application-optimized contrasts using different flip angle evolution” (SPACE) sequence allows better visualization of brain metastases (BMs) compared to “magnetization-prepared rapid acquisition gradient echo” (MPRAGE). We hypothesize that this better conspicuity leads to high-quality annotation (HAQ), enhancing deep learning (DL) algorithm detection of BMs on MPRAGE images.MethodsRetrospective contrast-enhanced (gadobutrol 0.1 mmol/kg) SPACE and MPRAGE data of 157 patients with BM were used, either annotated on MPRAGE resulting in normal annotation quality (NAQ) or on coregistered SPACE resulting in HAQ. Multiple DL methods were developed with NAQ or HAQ using either SPACE or MRPAGE images and evaluated on their detection performance using positive predictive value (PPV), sensitivity, and F1 score and on their delineation performance using volumetric Dice similarity coefficient, PPV, and sensitivity on one internal and four additional test datasets (660 patients).ResultsThe SPACE-HAQ model reached 0.978 PPV, 0.882 sensitivity, and 0.916 F1-score. The MPRAGE-HAQ reached 0.867, 0.839, and 0.840, the MPRAGE NAQ 0.964, 0.667, and 0.798, respectively (p ≥ 0.157). Relative to MPRAGE-NAQ, the MPRAGE-HAQ F1-score detection increased on all additional test datasets by 2.5–9.6 points (p < 0.016) and sensitivity improved on three datasets by 4.6–8.5 points (p < 0.001). Moreover, volumetric instance sensitivity improved by 3.6–7.6 points (p < 0.001).ConclusionHAQ improves DL methods without specialized imaging during application time. HAQ alone achieves about 40% of the performance improvements seen with SPACE images as input, allowing for fast and accurate, fully automated detection of small (< 1 cm) BMs.Relevance statementTraining with higher-quality annotations, created using the SPACE sequence, improves the detection and delineation sensitivity of DL methods for the detection of brain metastases (BMs)on MPRAGE images. This MRI cross-technique transfer learning is a promising way to increase diagnostic performance.Key PointsDelineating small BMs on SPACE MRI sequence results in higher quality annotations than on MPRAGE sequence due to enhanced conspicuity.Leveraging cross-technique ground truth annotations during training improved the accuracy of DL models in detecting and segmenting BMs.Cross-technique annotation may enhance DL models by integrating benefits from specialized, time-intensive MRI sequences while not relying on them.Further validation in prospective studies is needed.Graphical

PlaqueViT: a vision transformer model for fully automatic vessel and plaque segmentation in coronary computed tomography angiography.

To develop and evaluate a deep learning model for segmentation of the coronary artery vessels and coronary plaques in coronary computed tomography angiography (CCTA). CCTA image data from the Swedish CardioPulmonary BioImage Study (SCAPIS) was used for model development (n = 463 subjects) and testing (n = 123) and for an interobserver study (n = 65). A dataset from Linköping University Hospital (n = 28) was used for external validation. The model's ability to detect coronary artery disease (CAD) was tested in a separate SCAPIS dataset (n = 684). A deep ensemble (k = 6) of a customized 3D vision transformer model was used for voxelwise classification. The Dice coefficient, the average surface distance, Pearson's correlation coefficient, analysis of segmented volumes by intraclass correlation coefficient (ICC), and agreement (sensitivity and specificity) were used to analyze model performance. PlaqueViT segmented coronary plaques with a Dice coefficient = 0.55, an average surface distance = 0.98 mm and ICC = 0.93 versus an expert reader. In the interobserver study, PlaqueViT performed as well as the expert reader (Dice coefficient = 0.51 and 0.50, average surface distance = 1.31 and 1.15 mm, ICC = 0.97 and 0.98, respectively). PlaqueViT achieved 88% agreement (sensitivity 97%, specificity 76%) in detecting any coronary plaque in the test dataset (n = 123) and 89% agreement (sensitivity 95%, specificity 83%) in the CAD detection dataset (n = 684). We developed a deep learning model for fully automatic plaque detection and segmentation that identifies and delineates coronary plaques and the arterial lumen with similar performance as an experienced reader. Question A tool for fully automatic and voxelwise segmentation of coronary plaques in coronary CTA (CCTA) is important for both clinical and research usage of the CCTA examination. Findings Segmentation of coronary artery plaques by PlaqueViT was comparable to an expert reader's performance. Clinical relevance This novel, fully automatic deep learning model for voxelwise segmentation of coronary plaques in CCTA is highly relevant for large population studies such as the Swedish CardioPulmonary BioImage Study.

Universal slip detection of robotic hand with tactile sensing

Slip detection is to recognize whether an object remains stable during grasping, which can significantly enhance manipulation dexterity. In this study, we explore slip detection for five-finger robotic hands being capable of performing various grasp types, and detect slippage across all five fingers as a whole rather than concentrating on individual fingertips. First, we constructed a dataset collected during the grasping of common objects from daily life across six grasp types, comprising more than 200 k data points. Second, according to the principle of deep double descent, we designed a lightweight universal slip detection convolutional network for different grasp types (USDConvNet-DG) to classify grasp states (no-touch, slipping, and stable grasp). By combining frequency with time domain features, the network achieves a computation time of only 1.26 ms and an average accuracy of over 97% on both the validation and test datasets, demonstrating strong generalization capabilities. Furthermore, we validated the proposed USDConvNet-DG in real-time grasp force adjustment in real-world scenarios, showing that it can effectively improve the stability and reliability of robotic manipulation.

Real-time digital dermatitis detection in dairy cows on Android and iOS apps using computer vision techniques

Abstract The aim of the study was to deploy computer vision (CV) models for real-time detection of digital dermatitis (DD) lesions in cows using Android or iOS mobile applications. Early detection of DD lesions in dairy cows is crucial for prompt treatment and animal welfare. Android and iOS apps could facilitate routine and early DD detection in cows’ feet on dairy and beef farms. Upon detecting signs of DD, dairy farmers could implement preventive and treatment methods, including foot baths, topical treatment, hoof trimming, or quarantining cows affected by DD to prevent its spread. We applied transfer-learning to DD image data for 5 lesion classes, M0, M4H, M2, M2P, and M4P, on pretrained YOLOv5 model architecture using COCO-128 pretrained weights. The combination of localization loss, classification loss, and objectness loss was used for the optimization of prediction performance. The custom DD detection model was trained on 363 images of size 416 x 416 pixels and tested on 46 images. During model training, data were augmented to increase model robustness in different environments. The model was converted into TFLite format for Android devices and CoreML format for iOS devices. Techniques such as quantization were implemented to improve inference speed in real-world settings. The DD models achieved an average mAP (mean Average Precision) of 0.95 on the test dataset. When tested in real-time, iOS devices resulted in Cohen’s kappa value of 0.57 (95% CI: 0.49-0.65) averaged across the 5 lesion classes denoting the moderate agreement of the model detection with human investigators. The Android device resulted in a Cohen’s kappa value of 0.38 (95% CI: 0.29-0.47) denoting fair agreement between model and investigator. Combining M2 and M2P classes and M4H and M4P classes resulted in a Cohen’s kappa value of 0.65 (95% CI: 0.54-0.76) and 0.46 (95% CI: 0.35-0.57), for Android and iOS devices, respectively. For the two-class model (lesion vs. non-lesion), a Cohen’s kappa value of 0.74 (95% CI: 0.63-0.85) and 0.65 (95% CI: 0.52-0.78) was achieved for iOS and Android devices, respectively. iOS achieved a good inference time of 20 ms, compared to 57 ms on Android. Additionally, we deployed models on Ultralytics iOS and Android apps giving kappa scores of 0.56 (95% CI: 0.48-0.64) and 0.46 (95% CI: 0.37-0.55) respectively. Our custom iOS app surpassed the Ultralytics apps in terms of kappa score and the confidence score.

Artificial intelligence improves submucosal vessel detection during third space endoscopy.

While artificial intelligence (AI) shows high potential in decision support for diagnostic gastrointestinal endoscopy, its role in therapeutic endoscopy remains unclear. Third space endoscopic procedures pose the risk of intraprocedural bleeding. Therefore, we aimed to develop an AI algorithm for intraprocedural blood vessel detection. Using a test dataset with 101 standardized video clips containing 200 predefined submucosal blood vessels, 19 endoscopists were evaluated for the vessel detection rate (VDR) and time (VDT) with and without support of an AI algorithm. Test subjects were grouped according to experience in ESD. With AI support, endoscopists VDR increased from 56.4% [CI 54.1-58.6] to 72.4% [CI 70.3-74.4]. Endoscopists' VDT dropped from 6.7sec [CI 6.2-7.1] to 5.2sec [CI 4.8-5.7]. False positive (FP) readings appeared in 4.5% of frames and were marked significantly shorter than true positives (6.0sec [CI 5.28-6.70] vs. 0.7sec [CI 0.55-0.87]). AI improved the vessel detection rate and time of endoscopists during third space endoscopy. While these data need to be corroborated by clinical trials, AI may prove to be an invaluable tool for the improvement of endoscopic interventions.

Efficacy of Fine-Tuned Large Language Model in CT Protocol Assignment as Clinical Decision-Supporting System

Abstract Accurate CT protocol assignment is crucial for optimizing medical imaging procedures. The integration of large language models (LLMs) may be helpful, but its efficacy as a clinical decision support system for protocoling tasks remains unknown. This study aimed to develop and evaluate fine-tuned LLM specifically designed for CT protocoling, as well as assess its performance, both standalone and in concurrent use, in terms of effectiveness and efficiency within radiological workflows. This retrospective study included radiology tests for contrast-enhanced chest and abdominal CT examinations (2829/498/941 for training/validation/testing). Inputs involve the clinical indication section, age, and anatomic coverage. The LLM was fine-tuned for 15 epochs, selecting the best model by macro sensitivity in validation. Performance was then evaluated on 800 randomly selected cases from the test dataset. Two radiology residents and two radiologists assigned CT protocols with and without referencing the output of LLM to evaluate its efficacy as a clinical decision support system. The LLM exhibited high accuracy metrics, with top-1 and top-2 accuracies of 0.923 and 0.963, respectively, and a macro sensitivity of 0.907. It processed each case in an average of 0.39 s. The LLM, as a clinical decision support tool, improved accuracy both for residents (0.913 vs. 0.936) and radiologists (0.920 vs. 0.926 without and with LLM, respectively), with the improvement for residents being statistically significant (p = 0.02). Additionally, it reduced reading times by 14% for residents and 12% for radiologists. These results indicate the potential of LLMs to improve CT protocoling efficiency and diagnostic accuracy in radiological practice.

Feasibility and accuracy of single maxillary molar designed by an implicit neural network (INN)-based model: A comparative study.

This feasibility study aimed to develop and evaluate a novel implicit neural network (INN)-based model for the automatic reconstruction of biomimetic maxillary molar. A total of 500 sets of full dental scans containing intact right and left maxillary first molars (#3 and #14) and adjacent teeth were included in this study. The digital maxillary casts were duplicated: one set served as the original, while the other had one maxillary first molar removed. Two INN-based models were developed, including the point convolution for surface reconstruction (POCO-only) model and the POCO with point multilayer perceptron (POCO-PointMLP) model. Each model was trained with either 12,000 or 50,000 sampling points from the training dataset. Chamfer distance (CD), F-score, and Volumetric Intersection over Union (volumetric IoU) were used to evaluate the INN-based models and training outcomes. The best INN-based model was selected to generate 20 digital crown designs (GCs) automatically from randomly selected test datasets and compared with the digital crowns designed by experienced human laboratory technicians (TCs). Both GCs and TCs were compared with the original clinical crowns (OCs) using rootmean square(RMS) and color maps. Comparisons of RMS values were analyzed utilizing paired t-tests with statistical significance set at α=0.05. The performance of 2 INN-based models and training outcomes were evaluated and POCO-PointMLP trained with 50,000 sample points showed the best outcomes, with CD of 0.000119403, F-score of 0.919740566, and volumetric IoU of 0.943382978. This INN-based model was then selected to generate 20 digital crown designs (GCs) and compared with TCs for the RMS outcomes. The GCs exhibited numerically lower mean RMS compared to TCs without significant difference (0.2839±0.0307 versus 0.3026±0.0587mm, p=0.202). GCs and TCs both closely matched to the original clinical crowns (OCs). This study demonstrated the feasibility and accuracy of the proposed INN-based POCO-PointMLP model for automating the digital crown designs of maxillary first molars. The INN-based model-generated digital crown designs showed comparable 3D deviations to the ones designed by experienced human laboratory technicians, presenting a novel and efficient artificial neural network architecture for tooth digital design tasks with the expansion of training datasets, refined training hyperparameters, and integrated comprehensive assessments.

I-Brainer: Artificial intelligence/Internet of Things (AI/IoT)-Powered Detection of Brain Cancer.

Brain tumour is characterized by its aggressive nature and low survival rate and thus regarded as one of the deadliest diseases. Thus, miss-diagnosis or miss-classification of brain tumour can lead to miss treatment or incorrect treatment and reduce survival chances. Therefore, there is need to develop a technique that can identify and detect brain tumour at early stages. Here, we proposed a framework titled I-Brainer which is an Artificial Intelligence/Internet of Things (AI/IoT)-powered classification of MRI. We employed a Br35H+SARTAJ brain MRI dataset which contain 7023 total images which include No tumour, pituitary, meningioma and glioma. In order to accurately classified MRI into 4-class, we developed LeNet model from scratch, implemented 2 pretrained models which include EfficientNet and ResNet-50 as well feature extraction of these models coupled with 2 Machine Learning classifiers k-Nearest Neighbours (KNN) and Support Vector Machines (SVM). Evaluation and comparison of the performance of 3 models has shown that EfficientNet+SVM achieved the best result in terms of AUC (99%) and ResNet-50-KNN ranked higher in terms of accuracy (94%) on testing dataset. This framework can be harness by patients residing in remote areas and as confirmatory approach for medical experts.

Patient- and fraction-specific magnetic resonance volume reconstruction from orthogonal images with generative adversarial networks.

Although deep learning (DL) methods for reconstructing 3D magnetic resonance (MR) volumes from 2D MR images yield promising results, they require large amounts of training data to perform effectively. To overcome this challenge, fine-tuning-a transfer learning technique particularly effective for small datasets-presents a robust solution for developing personalized DL models. A 2D to 3D conditional generative adversarial network (GAN) model with a patient- and fraction-specific fine-tuning workflow was developed to reconstruct synthetic 3D MR volumes using orthogonal 2D MR images for online dose adaptation. A total of 2473 3D MR volumes were collected from 43 patients. The training and test datasets were separated into 34 and 9 patients, respectively. All patients underwent MR-guided adaptive radiotherapy using the same imaging protocol. The population data contained 2047 3D MR volumes from the training dataset. Population data were used to train the population-based GAN model. For each fraction of the remaining patients, the population model was fine-tuned with the 3D MR volumes acquired before beam irradiation of the fraction, named the fine-tuned model. The performance of the fine-tuned model was tested using the 3D MR volume acquired immediately after the beam delivery of the fraction. The model's input was a pair of axial and sagittal MR images at the isocenter level, and the output was a 3D MR volume. Model performance was evaluated using the structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), root mean square error (RMSE), and mean absolute error (MAE). Moreover, the prostate, bladder, and rectum in the predicted MR images were manually segmented. To assess geometric accuracy, the 2D Dice Similarity Coefficient (DSC) and 2D Hausdorff Distance (HD) were calculated. A total of 84 3D MR volumes were included in the performance testing. The mean±standard deviation (SD) of SSIM, PSNR, RMSE, and MAE were 0.64±0.10, 93.9±1.5dB, 0.050±0.009, and 0.036±0.007 for the population model and 0.72±0.09, 96.2±1.8dB, 0.041±0.007, and 0.028±0.006 for the fine-tuned model, respectively. The image quality of the fine-tuned model was significantly better than that of the population model (p<0.05). The mean±SD of DSC and HD of the population model were 0.79±0.08 and 1.70±2.35mm for prostate, 0.81±0.10 and 2.75±1.53mm for bladder, and 0.72±0.08 and 1.93±0.59mm for rectum. Contrarily, the mean±SD of DSC and HD of the fine-tuned model were 0.83±0.06 and 1.29±0.77mm for prostate, 0.85±0.07 and 2.16±1.09mm for bladder, and 0.77±0.08 and 1.57±0.52mm for rectum. The geometric accuracy of the fine-tuned model was significantly improved than that of the population model (p<0.05). By employing a patient- and fraction-specific fine-tuning approach, the GAN model demonstrated promising accuracy despite limited data availability.

Deep plug-and-play HIO approach for phase retrieval

In the phase retrieval problem, the aim is the recovery of an unknown image from intensity-only measurements such as Fourier intensity. Although there are several solution approaches, solving this problem is challenging due to its nonlinear and ill-posed nature. Recently, learning-based approaches have emerged as powerful alternatives to the analytical methods for several inverse problems. In the context of phase retrieval, a novel plug-and-play approach, to our knowledge, that exploits learning-based prior and efficient update steps has been presented at the Computational Optical Sensing and Imaging topical meeting, with demonstrated state-of-the-art performance. The key idea was to incorporate learning-based prior to the Gerchberg-Saxton type algorithms through plug-and-play regularization. In this paper, we present the mathematical development of the method including the derivation of its analytical update steps based on half-quadratic splitting and comparatively evaluate its performance through extensive simulations on a large test dataset. The results show the effectiveness of the method in terms of image quality, computational efficiency, and robustness to initialization and noise.