External Test Dataset Research Articles

A human-in-the-loop based Bayesian network approach to improve imbalanced radiation outcomes prediction for hepatocellular cancer patients with stereotactic body radiotherapy.

Imbalanced outcome is one of common characteristics of oncology datasets. Current machine learning approaches have limitation in learning from such datasets. Here, we propose to resolve this problem by utilizing a human-in-the-loop (HITL) approach, which we hypothesize will also lead to more accurate and explainable outcome prediction models. A total of 119 HCC patients with 163 tumors were used in the study. 81 patients with 104 tumors from the University of Michigan Hospital treated with SBRT were considered as a discovery dataset for radiation outcomes model building. The external testing dataset included 59 tumors from 38 patients with SBRT from Princess Margaret Hospital. In the discovery dataset, 100 tumors from 77 patients had local control (LC) (96% of 104 tumors) and 23 patients had at least one grade increment of ALBI (I-ALBI) during six-month follow up (28% of 81 patients). Each patient had a total of 110 features, where 15 or 20 features were identified by physicians as expert knowledge features (EKFs) for LC or I-ALBI prediction. We proposed a HITL based Bayesian network (HITL-BN) approach to enhance the capability of selecting important features from imbalanced data in terms of accuracy and explainability through humans' participation by integrating feature importance ranking and Markov blanket algorithms. A pure data-driven Bayesian network (PD-BN) method was applied to the same discovery dataset of HCC patients as a benchmark. In the training and testing phases, the areas under receiver operating characteristic curves of the HITL-BN models for LC or I-ALBI prediction during SBRT are 0.85 (95% confidence interval: 0.75-0.95) or 0.89 (0.81-0.95) and 0.77 or 0.78, respectively. They significantly outperformed the during-treatment PD-BN model in predicting LC or I-ALBI based on the discovery cross-validation and testing datasets from the Delong tests. By allowing the human expert to be part of the model building process, the HITL-BN approach yielded significantly improved accuracy as well as better explainability when dealing with imbalanced outcomes in the prediction of post-SBRT treatment response of HCC patients when compared to the PD-BN method.

Deciphering clinical abbreviations with a privacy protecting machine learning system

Physicians write clinical notes with abbreviations and shorthand that are difficult to decipher. Abbreviations can be clinical jargon (writing “HIT” for “heparin induced thrombocytopenia”), ambiguous terms that require expertise to disambiguate (using “MS” for “multiple sclerosis” or “mental status”), or domain-specific vernacular (“cb” for “complicated by”). Here we train machine learning models on public web data to decode such text by replacing abbreviations with their meanings. We report a single translation model that simultaneously detects and expands thousands of abbreviations in real clinical notes with accuracies ranging from 92.1%-97.1% on multiple external test datasets. The model equals or exceeds the performance of board-certified physicians (97.6% vs 88.7% total accuracy). Our results demonstrate a general method to contextually decipher abbreviations and shorthand that is built without any privacy-compromising data.

Prediction Models for Suicide Attempts among Adolescents Using Machine Learning Techniques.

Suicide attempts (SAs) in adolescents are difficult to predict although it is a leading cause of death among adolescents. This study aimed to develop and evaluate SA prediction models based on six different machine learning (ML) algorithms for Korean adolescents using data from online surveys. Data were extracted from the 2011-2018 Korea Youth Risk Behavior Survey (KYRBS), an ongoing annual national survey. The participants comprised 468,482 nationally representative adolescents from 400 middle and 400 high schools, aged 12 to 18. The models were trained using several classic ML methods and then tested on internal and external independent datasets; performance metrics were calculated. Data analysis was performed from March 2020 to June 2020. Among the 468,482 adolescents included in the analysis, 15,012 cases (3.2%) were identified as having made an SA. Three features (suicidal ideation, suicide planning, and grade) were identified as the most important predictors. The performance of the six ML models on the internal testing dataset was good, with both the area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC) ranging from 0.92 to 0.94. Although the AUROC of all models on the external testing dataset (2018 KYRBS) ranged from 0.93 to 0.95, the AUPRC of the models was approximately 0.5. The developed and validated SA prediction models can be applied to detect high risks of SA. This approach could facilitate early intervention in the suicide crisis and may ultimately contribute to suicide prevention for adolescents.

Construction of the model for predicting prognosis by key genes regulating EGFR-TKI resistance.

Background: Previous studies have suggested that patients with lung adenocarcinoma (LUAD) will significantly benefit from epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKI). However, many LUAD patients will develop resistance to EGFR-TKI. Thus, our study aims to develop models to predict EGFR-TKI resistance and the LUAD prognosis. Methods: Two Gene Expression Omnibus (GEO) datasets (GSE31625 and GSE34228) were used as the discovery datasets to find the common differentially expressed genes (DEGs) in EGFR-TKI resistant LUAD profiles. The association of these common DEGs with LUAD prognosis was investigated in The Cancer Genome Atlas (TCGA) database. Moreover, we constructed the risk score for prognosis prediction of LUAD by LASSO analysis. The performance of the risk score for predicting LUAD prognosis was calculated using an independent dataset (GSE37745). A random forest model by risk score genes was trained in the training dataset, and the diagnostic ability for distinguishing sensitive and EGFR-TKI resistant samples was validated in the internal testing dataset and external testing datasets (GSE122005, GSE80344, and GSE123066). Results: From the discovery datasets, 267 common upregulated genes and 374 common downregulated genes were identified. Among these common DEGs, there were 59 genes negatively associated with prognosis, while 21 genes exhibited positive correlations with prognosis. Eight genes (ABCC2, ARL2BP, DKK1, FUT1, LRFN4, PYGL, SMNDC1, and SNAI2) were selected to construct the risk score signature. In both the discovery and independent validation datasets, LUAD patients with the higher risk score had a poorer prognosis. The nomogram based on risk score showed good performance in prognosis prediction with a C-index of 0.77. The expression levels of ABCC2, ARL2BP, DKK1, LRFN4, PYGL, SMNDC1, and SNAI2 were positively related to the resistance of EGFR-TKI. However, the expression level of FUT1 was favorably correlated with EGFR-TKI responsiveness. The RF model worked wonderfully for distinguishing sensitive and resistant EGFR-TKI samples in the internal and external testing datasets, with predictive area under the curves (AUC) of 0.973 and 0.817, respectively. Conclusion: Our investigation revealed eight genes associated with EGFR-TKI resistance and provided models for EGFR-TKI resistance and prognosis prediction in LUAD patients.

Deep Learning Models to Predict Fatal Pneumonia Using Chest X-Ray Images.

Chest X-ray (CXR) is indispensable to the assessment of severity, diagnosis, and management of pneumonia. Deep learning is an artificial intelligence (AI) technology that has been applied to the interpretation of medical images. This study investigated the feasibility of classifying fatal pneumonia based on CXR images using deep learning models on publicly available platforms. CXR images of patients with pneumonia at diagnosis were labeled as fatal or nonfatal based on medical records. We applied CXR images from 1031 patients with nonfatal pneumonia and 243 patients with fatal pneumonia for training and self-evaluation of the deep learning models. All labeled CXR images were randomly allocated to the training, validation, and test datasets of deep learning models. Data augmentation techniques were not used in this study. We created two deep learning models using two publicly available platforms. The first model showed an area under the precision-recall curve of 0.929 with a sensitivity of 50.0% and a specificity of 92.4% for classifying fatal pneumonia. We evaluated the performance of our deep learning models using sensitivity, specificity, PPV, negative predictive value (NPV), accuracy, and F1 score. Using the external validation test dataset of 100 CXR images, the sensitivity, specificity, accuracy, and F1 score were 68.0%, 86.0%, 77.0%, and 74.7%, respectively. In the original dataset, the performance of the second model showed a sensitivity, specificity, and accuracy of 39.6%, 92.8%, and 82.7%, respectively, while external validation showed values of 38.0%, 92.0%, and 65.0%, respectively. The F1 score was 52.1%. These results were comparable to those obtained by respiratory physicians and residents. The deep learning models yielded good accuracy in classifying fatal pneumonia. By further improving the performance, AI could assist physicians in the severity assessment of patients with pneumonia.

Automatic diagnosis and grading of Prostate Cancer with weakly supervised learning on whole slide images

Background:The workflow of prostate cancer diagnosis and grading is cumbersome and the results suffer from substantial inter-observer variability. Recent trials have shown potential in using machine learning to develop automated systems to address this challenge. Most automated deep learning systems for prostate cancer Gleason grading focused on supervised learning requiring demanding fine-grained pixel-level annotations. Methods:A weakly-supervised deep learning model with slide-level labels is presented in this study for the diagnosis and grading of prostate cancer with whole slide image (WSI). WSIs are first cropped into small patches and then processed with a deep learning model to extract patch-level features. A graph convolution network (GCN) is used to aggregate the features for classifications. Throughout the training process, the noisy labels are progressively filtered out to reduce inter-observer variations in clinical reports. Finally, multi-center independent test cohorts with 6,174 slides are collected to evaluate the prostate cancer diagnosis and grading performance of our model. Results:The cancer diagnosis (2-level classification) results on two external test sets (n= 4,675, n= 844) show an area under the receiver operating characteristic curve (AUC) of 0.985 and 0.986. The Gleason grading (6-level classification) results reach 0.931 quadratic weighted kappa on the internal test set (n= 531). It generalizes well on the external test dataset (n= 844) with 0.801 quadratic weighted kappa with the reference standard set independently. The model enables pathological meaningful interpretability by visualizing the most attended lesions which are highly consistent with expert annotations. Conclusion:The proposed model incorporates a graph network in weakly supervised learning with only slide-level reports. A robust learning strategy is also employed to correct the label noise. It is highly accurate (>0.985 AUC for diagnosis) and also interpretable with intuitive heatmap visualization. It can be unified with a digital pathology pipeline to deliver prostate cancer metrics for a pathology report.

NIMG-83. MULTIMODAL PET/MRI RADIOMICS AND CLINICAL PARAMETERS FOR OVERALL SURVIVAL PREDICTION IN PATIENTS WITH IDH WILDTYPE GLIOBLASTOMA

Abstract BACKGROUND Currently, most radiomics studies on survival prediction in brain tumor patients are based on MRI only. The goal of our study was to evaluate multimodal radiomics derived from amino acid PET/MRI and clinical parameters for survival prediction in patients with newly diagnosed IDH wildtype glioblastoma. METHODS Sixty-three patients with newly diagnosed IDH wildtype glioblastoma were evaluated retrospectively. At initial diagnosis, all patients underwent structural MRI and O-(2-[18F]fluoroethyl)-L-tyrosine (FET) PET. Tumor volumes were automatically segmented using a deep learning-based tool followed by visual inspection. Predefined and deep radiomics features were extracted from both imaging modalities. Feature repeatability analyses and feature selection were performed to avoid overfitting. Cox regression models for overall survival were built from clinical parameters such as age or the extent of resection, radiomics features, and combinations thereof, and finally validated using 5-fold cross-validation. Further evaluation of the model in an external test dataset is ongoing. RESULTS The median overall survival was 12 months (range, 0-64 months). Higher age and larger FET PET tumor volumes were significantly correlated with shorter overall survival (age, r=-0.39, p< 0.001; volume, r=-0.31, p< 0.05). Models solely based on predefined FET PET or MRI radiomics features showed a similar mean concordance index (C-index) as the model based on clinical parameters (C-indices, 0.68±0.04; 0.64±0.03; and 0.69±0.08, respectively). Multimodal radiomics based on predefined and deep features yielded improved C-indices of 0.75±0.06 and 0.72±0.09, respectively. A model based on multimodal radiomics and clinical parameters achieved the best prognostic performance (C-index, 0.80±0.04). CONCLUSION Our results suggest an added clinical value of multimodal FET PET/MRI radiomics with clinical parameters for the non-invasive survival prediction in patients with IDH wildtype glioblastoma.

A Radiomics Approach on Chest CT Distinguishes Primary Lung Cancer from Solitary Lung Metastasis in Colorectal Cancer Patients

This study utilized a radiomics approach combined with a machine learning algorithm to distinguish primary lung cancer (LC) from solitary lung metastasis (LM) in colorectal cancer (CRC) patients with a solitary pulmonary nodule (SPN). In a retrospective study, 239 patients who underwent chest computerized tomography (CT) at three different institutions between 2011 and 2019 and were diagnosed as primary LC or solitary LM were included. The data from the first institution were divided into training and internal testing datasets. The data from the second and third institutions were used as an external testing dataset. Radiomic features were extracted from the intra and perinodular regions of interest (ROI). After a feature selection process, Support vector machine (SVM) was used to train models for classifying between LC and LM. The performances of the SVM classifiers were evaluated with both the internal and external testing datasets. The performances of the model were compared to those of two radiologists who reviewed the CT images of the testing datasets for the binary prediction of LC versus LM. The SVM classifier trained with the radiomic features from the intranodular ROI and achieved the sensitivity/specificity of 0.545/0.828 in the internal test dataset, and 0.833/0.964 in the external test dataset, respectively. The SVM classifier trained with the combined radiomic features from the intra- and perinodular ROIs achieved the sensitivity/specificity of 0.545/0.966 in the internal test dataset, and 0.833/1.000 in the external test data set, respectively. Two radiologists demonstrated the sensitivity/specificity of 0.545/0.966 and 0.636/0.828 in the internal test dataset, and 0.917/0.929 and 0.833/0.929 in the external test dataset, which were comparable to the performance of the model trained with the combined radiomics features. Our results suggested that the machine learning classifiers trained using radiomics features of SPN in CRC patients can be used to distinguish the primary LC and the solitary LM with a similar level of performance to radiologists.

Covid-19 detection using chest X-rays: is lung segmentation important for generalization?

PurposeWe evaluated the generalization capability of deep neural networks (DNNs) in the task of classifying chest X-rays as Covid-19, normal or pneumonia, when trained in a relatively small and mixed datasets.MethodsWe proposed a DNN to perform lung segmentation and classification, stacking a segmentation module (U-Net), an original intermediate module and a classification module (DenseNet201). To evaluate generalization capability, we tested the network with an external dataset (from distinct localities) and used Bayesian inference to estimate the probability distributions of performance metrics. Furthermore, we introduce a novel evaluation technique, which uses layer-wise relevance propagation (LRP) and Brixia scores to compare the DNN grounds for decision with radiologists.ResultsThe proposed DNN achieved 0.917 AUC (area under the ROC curve) on the external test dataset, surpassing a DenseNet without segmentation, which showed 0.906 AUC. Bayesian inference indicated mean accuracy of 76.1% and [0.695, 0.826] 95% HDI (high-density interval, which concentrates 95% of the metric’s probability mass) with segmentation and, without segmentation, 71.7% and [0.646, 0.786].ConclusionEmploying an analysis based on LRP and Brixia scores, we discovered that areas where radiologists found strong Covid-19 symptoms are the most important for the stacked DNN classification. External validation showed smaller accuracies than internal, indicating difficulty in generalization, which is positively affected by lung segmentation. Finally, the performance on the external dataset and the analysis with LRP suggest that DNNs can successfully detect Covid-19 even when trained on small and mixed datasets.

A Deep Learning Segmentation Pipeline for Cardiac T1 Mapping Using MRI Relaxation-based Synthetic Contrast Augmentation.

To design and evaluate an automated deep learning method for segmentation and analysis of cardiac MRI T1 maps with use of synthetic T1-weighted images for MRI relaxation-based contrast augmentation. This retrospective study included MRI scans acquired between 2016 and 2019 from 100 patients (mean age ± SD, 55 years ± 13; 72 men) across various clinical abnormalities with use of a modified Look-Locker inversion recovery, or MOLLI, sequence to quantify native T1 (T1native), postcontrast T1 (T1post), and extracellular volume (ECV). Data were divided into training (n = 60) and internal (n = 40) test subsets. "Synthetic" T1-weighted images were generated from the T1 exponential inversion-recovery signal model at a range of optimal inversion times, yielding high blood-myocardium contrast, and were used for contrast-based image augmentation during training and testing of a convolutional neural network for myocardial segmentation. Automated segmentation, T1, and ECV were compared with experts with use of Dice similarity coefficients (DSCs), correlation coefficients, and Bland-Altman analysis. An external test dataset (n = 147) was used to assess model generalization. Internal testing showed high myocardial DSC relative to experts (0.81 ± 0.08), which was similar to interobserver DSC (0.81 ± 0.08). Automated segmental measurements strongly correlated with experts (T1native, R = 0.87; T1post, R = 0.91; ECV, R = 0.92), which were similar to interobserver correlation (T1native, R = 0.86; T1post, R = 0.94; ECV, R = 0.95). External testing showed strong DSC (0.80 ± 0.09) and T1native correlation (R = 0.88) between automatic and expert analysis. This deep learning method leveraging synthetic contrast augmentation may provide accurate automated T1 and ECV analysis for cardiac MRI data acquired across different abnormalities, centers, scanners, and T1 sequences.Keywords: MRI, Cardiac, Tissue Characterization, Segmentation, Convolutional Neural Network, Deep Learning Algorithms, Machine Learning Algorithms, Supervised Learning Supplemental material is available for this article. © RSNA, 2022.

Reliable prediction of crystal density of high nitrogen-containing organic compounds as powerful, less sensitive, eco-friendly energetic materials for dependable assessment of their performance

Many efforts have been done in recent years to seek more powerful, less sensitive, eco-friendly high nitrogen-containing organic compounds (HNCOCs). It is essential to have reliable methods for the prediction of crystal density of the new proposed HNCOCs because increasing density is associated with increased performance. This work uses structural parameters of HNCOCs and semiempirical quantum mechanical calculations using the PM6 method for reliable prediction of their densities at or near room temperature. A core correlation is introduced for the calculation of the density of HNCOCs based on the number of functional groups –NH2, >NH, –OH, and –N3 as well as the calculated density from the PM6 method. The reliability of the core correlation is improved by considering two non-additive correcting functions, which can increase or decrease the predicted values of core density. The reported densities of 137 HNCOCs as training (117 compounds) and test (20 chemicals) sets are used to establish the new model. High reliability of the core and improved models are compared with the computed results of the best available molecular surface electrostatic potential (MESP)-based method. The values of Root Mean Square Error (RMSE) for the improved model and comparative MESP-based method are 0.0472 and 0.0951 g/cm3, respectively. High reliability of the new method was also confirmed as compared to four MESP-based models for 32 compounds of external test data set where RMSE of the improved model (0.0674 g/cm3) is smaller than four MSPE-based methods, i.e. 0.126, 0.276, 0.0716, and 0.0703 g/cm3.

Automated deep neural network analysis of lateral cephalogram data can aid in detecting obstructive sleep apnea.

Information on obstructive sleep apnea (OSA) is often latently detected in diagnostic tests conducted for other purposes, providing opportunities for maximizing value. This study aimed to develop a convolutional neural network (CNN) to identify the risk of OSA using lateral cephalograms. The lateral cephalograms of 5,648 individuals (mean age, 49.0 ± 15.8 years; men, 62.3%) with or without OSA were collected and divided into training, validation, and internal test datasets in a 5:2:3 ratio. A separate external test dataset (n = 378) was used. A densely connected CNN was trained to diagnose OSA using a cephalogram. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC). Gradient-weighted class activation mapping (Grad-CAM) was used to evaluate the region of focus, and the relationships between the model outputs, anthropometric characteristics, and OSA severity were evaluated. The AUROC of the model for the presence of OSA was 0.82 (95% confidence interval, 0.80-0.84) and 0.73 (95% confidence interval, 0.65-0.81) in the internal and external test datasets, respectively. Grad-CAM demonstrated that the model focused on the area of the tongue base and oropharynx in the cephalogram. Sigmoid output values were positively correlated with OSA severity, body mass index, and neck and waist circumference. Deep learning may help develop a model that classifies OSA using a cephalogram, which may be clinically useful in the appropriate context. The definition of ground truth was the main limitation of this study. Jeong H-G, Kim T, Hong JE, etal. Automated deep neural network analysis of lateral cephalogram data can aid in detecting obstructive sleep apnea. J Clin Sleep Med. 2023;19(2):327-337.

TransCluster: A Cell-Type Identification Method for single-cell RNA-Seq data using deep learning based on transformer.

Recent advances in single-cell RNA sequencing (scRNA-seq) have accelerated the development of techniques to classify thousands of cells through transcriptome profiling. As more and more scRNA-seq data become available, supervised cell type classification methods using externally well-annotated source data become more popular than unsupervised clustering algorithms. However, accurate cellular annotation of single cell transcription data remains a significant challenge. Here, we propose a hybrid network structure called TransCluster, which uses linear discriminant analysis and a modified Transformer to enhance feature learning. It is a cell-type identification tool for single-cell transcriptomic maps. It shows high accuracy and robustness in many cell data sets of different human tissues. It is superior to other known methods in external test data set. To our knowledge, TransCluster is the first attempt to use Transformer for annotating cell types of scRNA-seq, which greatly improves the accuracy of cell-type identification.

Quantum-based machine learning and AI models to generate force field parameters for drug-like small molecules.

Force fields for drug-like small molecules play an essential role in molecular dynamics simulations and binding free energy calculations. In particular, the accurate generation of partial charges on small molecules is critical to understanding the interactions between proteins and drug-like molecules. However, it is a time-consuming process. Thus, we generated a force field for small molecules and employed a machine learning (ML) model to rapidly predict partial charges on molecules in less than a minute of time. We performed density functional theory (DFT) calculation for 31770 small molecules that covered the chemical space of drug-like molecules. The partial charges for the atoms in a molecule were predicted using an ML model trained on DFT-based atomic charges. The predicted values were comparable to the charges obtained from DFT calculations. The ML model showed high accuracy in the prediction of atomic charges for external test data sets. We also developed neural network (NN) models to assign atom types, phase angles and periodicities. All the models performed with high accuracy on test data sets. Our code calculated all the descriptors that were needed for the prediction of force field parameters and produced topologies for small molecules by combining results from ML and NN models. To assess the accuracy of the predicted force field parameters, we calculated solvation free energies for small molecules, and the results were in close agreement with experimental free energies. The AI-generated force field was effective in the fast and accurate generation of partial charges and other force field parameters for small drug-like molecules.

RATING: Medical knowledge-guided rheumatoid arthritis assessment from multimodal ultrasound images via deep learning

SummaryMultimodal ultrasound has demonstrated its power in the clinical assessment of rheumatoid arthritis (RA). However, for radiologists, it requires strong experience. In this paper, we propose a rheumatoid arthritis knowledge guided (RATING) system that automatically scores the RA activity and generates interpretable features to assist radiologists' decision-making based on deep learning. RATING leverages the complementary advantages of multimodal ultrasound images and solves the limited training data problem with self-supervised pretraining. RATING outperforms all of the existing methods, achieving an accuracy of 86.1% on a prospective test dataset and 85.0% on an external test dataset. A reader study demonstrates that the RATING system improves the average accuracy of 10 radiologists from 41.4% to 64.0%. As an assistive tool, not only can RATING indicate the possible lesions and enhance the diagnostic performance with multimodal ultrasound but it can also enlighten the road to human-machine collaboration in healthcare.

Rapid Assessment of Microbial Quality in Edible Seaweeds Using Sensor Techniques Based on Spectroscopy, Imaging Analysis and Sensors Mimicking Human Senses.

The expansion of the seaweed aquaculture sector along with the rapid deterioration of these products escalates the importance of implementing rapid, real-time techniques for their quality assessment. Seaweed samples originating from Scotland and Ireland were stored under various temperature conditions for specific time intervals. Microbiological analysis was performed throughout storage to assess the total viable counts (TVC), while in parallel FT-IR spectroscopy, multispectral imaging (MSI) and electronic nose (e-nose) analyses were conducted. Machine learning models (partial least square regression (PLS-R)) were developed to assess any correlations between sensor and microbiological data. Microbial counts ranged from 1.8 to 9.5 log CFU/g, while the microbial growth rate was affected by origin, harvest year and storage temperature. The models developed using FT-IR data indicated a good prediction performance on the external test dataset. The model developed by combining data from both origins resulted in satisfactory prediction performance, exhibiting enhanced robustness from being origin unaware towards microbiological population prediction. The results of the model developed with the MSI data indicated a relatively good prediction performance on the external test dataset in spite of the high RMSE values, whereas while using e-nose data from both MI and SAMS, a poor prediction performance of the model was reported.

LASSO Regression-Based Diagnosis of Acute ST-Segment Elevation Myocardial Infarction (STEMI) on Electrocardiogram (ECG).

Electrocardiogram (ECG) is an important tool for the detection of acute ST-segment elevation myocardial infarction (STEMI). However, machine learning (ML) for the diagnosis of STEMI complicated with arrhythmia and infarct-related arteries is still underdeveloped based on real-world data. Therefore, we aimed to develop an ML model using the Least Absolute Shrinkage and Selection Operator (LASSO) to automatically diagnose acute STEMI based on ECG features. A total of 318 patients with STEMI and 502 control subjects were enrolled from Jan 2017 to Jun 2019. Coronary angiography was performed. A total of 180 automatic ECG features of 12-lead ECG were input into the model. The LASSO regression model was trained and validated by the internal training dataset and tested by the internal and external testing datasets. A comparative test was performed between the LASSO regression model and different levels of doctors. To identify the STEMI and non-STEMI, the LASSO model retained 14 variables with AUCs of 0.94 and 0.93 in the internal and external testing datasets, respectively. The performance of LASSO regression was similar to that of experienced cardiologists (AUC: 0.92) but superior (p < 0.05) to internal medicine residents, medical interns, and emergency physicians. Furthermore, in terms of identifying left anterior descending (LAD) or non-LAD, LASSO regression achieved AUCs of 0.92 and 0.98 in the internal and external testing datasets, respectively. This LASSO regression model can achieve high accuracy in diagnosing STEMI and LAD vessel disease, thus providing an assisting diagnostic tool based on ECG, which may improve the early diagnosis of STEMI.

Evaluation of Generative Adversarial Networks for High-Resolution Synthetic Image Generation of Circumpapillary Optical Coherence Tomography Images for Glaucoma

Deep learning (DL) networks require large data sets for training, which can be challenging to collect clinically. Generative models could be used to generate large numbers of synthetic optical coherence tomography (OCT) images to train such DL networks for glaucoma detection. To assess whether generative models can synthesize circumpapillary optic nerve head OCT images of normal and glaucomatous eyes and determine the usability of synthetic images for training DL models for glaucoma detection. Progressively growing generative adversarial network models were trained to generate circumpapillary OCT scans. Image gradeability and authenticity were evaluated on a clinical set of 100 real and 100 synthetic images by 2 clinical experts. DL networks for glaucoma detection were trained with real or synthetic images and evaluated on independent internal and external test data sets of 140 and 300 real images, respectively. Evaluations of the clinical set between the experts were compared. Glaucoma detection performance of the DL networks was assessed using area under the curve (AUC) analysis. Class activation maps provided visualizations of the regions contributing to the respective classifications. A total of 990 normal and 862 glaucomatous eyes were analyzed. Evaluations of the clinical set were similar for gradeability (expert 1: 92.0%; expert 2: 93.0%) and authenticity (expert 1: 51.8%; expert 2: 51.3%). The best-performing DL network trained on synthetic images had AUC scores of 0.97 (95% CI, 0.95-0.99) on the internal test data set and 0.90 (95% CI, 0.87-0.93) on the external test data set, compared with AUCs of 0.96 (95% CI, 0.94-0.99) on the internal test data set and 0.84 (95% CI, 0.80-0.87) on the external test data set for the network trained with real images. An increase in the AUC for the synthetic DL network was observed with the use of larger synthetic data set sizes. Class activation maps showed that the regions of the synthetic images contributing to glaucoma detection were generally similar to that of real images. DL networks trained with synthetic OCT images for glaucoma detection were comparable with networks trained with real images. These results suggest potential use of generative models in the training of DL networks and as a means of data sharing across institutions without patient information confidentiality issues.

Automated Lung Segmentation from Computed Tomography Images of Normal and COVID-19 Pneumonia Patients.

Background: Automated image segmentation is an essential step in quantitative image analysis. This study assesses the performance of a deep learning-based model for lung segmentation from computed tomography (CT) images of normal and COVID-19 patients.Methods: A descriptive-analytical study was conducted from December 2020 to April 2021 on the CT images of patients from various educational hospitals affiliated with Mashhad University of Medical Sciences (Mashhad, Iran). Of the selected images and corresponding lung masks of 1,200 confirmed COVID-19 patients, 1,080 were used to train a residual neural network. The performance of the residual network (ResNet) model was evaluated on two distinct external test datasets, namely the remaining 120 COVID-19 and 120 normal patients. Different evaluation metrics such as Dice similarity coefficient (DSC), mean absolute error (MAE), relative mean Hounsfield unit (HU) difference, and relative volume difference were calculated to assess the accuracy of the predicted lung masks. The Mann-Whitney U test was used to assess the difference between the corresponding values in the normal and COVID-19 patients. P<0.05 was considered statistically significant.Results: The ResNet model achieved a DSC of 0.980 and 0.971 and a relative mean HU difference of -2.679% and -4.403% for the normal and COVID-19 patients, respectively. Comparable performance in lung segmentation of normal and COVID-19 patients indicated the model’s accuracy for identifying lung tissue in the presence of COVID-19-associated infections. Although a slightly better performance was observed in normal patients.Conclusion: The ResNet model provides an accurate and reliable automated lung segmentation of COVID-19 infected lung tissue.A preprint version of this article was published on arXiv before formal peer review (https://arxiv.org/abs/2104.02042).

Metabolite biomarker discovery for human gastric cancer using dried blood spot mass spectrometry metabolomic approach

As one of the most common malignancies, gastric cancer (GC) is the third leading cause of cancer-related deaths in China. GC is asymptomatic in early stages, and the majority of GC mortality is due to delayed symptoms. It is an urgent task to find reliable biomarkers for the identification of GC in order to improve outcomes. A combination of dried blood spot sampling and direct infusion mass spectrometry (MS) technology was used to measure blood metabolic profiles for 166 patients with GC and 183 healthy individuals, and 93 metabolites including amino acids, carnitine/acylcarnitines and their derivatives, and related ratios were quantified. Multiple algorithms were used to characterize the changes of metabolic profiles in patients with GC compared to healthy individuals. A biomarker panel was identified in training set, and assessed by tenfold cross-validation and external test data set. After systematic selection of 93 metabolites, a biomarker panel consisting of Ala, Arg, Gly, Orn, Tyr/Cit, Val/Phe, C4-OH, C5/C3, C10:2 shows the potential to distinguish patients with GC from healthy individuals in tenfold cross-validation model (sensitivity: 0.8750, specificity: 0.9006) and test set (sensitivity: 0.9545, specificity: 0.8636). This metabolomic analysis makes contribution to the identification of disease-associated biomarkers and to the development of new diagnostic tools for patients with GC.