External Test Research Articles

An Integrated Nomogram Combining Deep Learning and Radiomics for Predicting Malignancy of Pulmonary Nodules Using CT-Derived Nodules and Adipose Tissue: AMulticenter Study.

Correctly distinguishing between benign and malignant pulmonary nodules can avoid unnecessary invasive procedures. This study aimed to construct a deep learning radiomics clinical nomogram (DLRCN) for predicting malignancy of pulmonary nodules. One thousand and ninety-eight patients with 6-30 mm pulmonary nodules who received histopathologic diagnosis at 3 centers were included and divided into a primary cohort (PC), an internal test cohort (I-T), and two external test cohorts (E-T1, E-T2). The DLRCN was built by integrating adipose tissue radiomics features, intranodular and perinodular deep learning features, and clinical characteristics for diagnosing malignancy of pulmonary nodules. The least absolute shrinkage and selection operator (LASSO) was used for feature selection. The performance of DLRCN was assessed with respect to its calibration curve, area under the curve (AUC), and decision curve analysis (DCA). Furthermore, we compared it with three radiologists. The net reclassification improvement (NRI), integrated discrimination improvement (IDI), and subgroup analysis were also taken into account. The incorporation of adipose tissue radiomics features led to significant NRI and IDI (NRI = 1.028, p < 0.05, IDI = 0.137, p < 0.05). In the I-T, E-T1, and E-T2, the AUCs of DLRCN were 0.946 (95% CI: 0.936, 0.955), 0.948 (95% CI: 0.933, 0.963) and 0.962 (95% CI: 0.945, 0.979), The calibration curve revealed good predictive accuracy between the actual probability and predicted probability (p > 0.05). DCA showed that the DLRCN was clinically useful. Under equal specificity, the sensitivity of DLRCN increased by 8.6% compared to radiologist assessments. The subgroup analysis conducted on adipose tissue radiomics features further demonstrated their supplementary value in determining the malignancy of pulmonary nodules. The DLRCN demonstrated good performance in predicting the malignancy of pulmonary nodules, which was comparable to radiologist assessments. The adipose tissue radiomics features have notably enhanced the performance of DLRCN.

Deep Learning and Single-Cell Sequencing Analyses Unveiling Key Molecular Features in the Progression of Carotid Atherosclerotic Plaque.

Rupture of advanced carotid atherosclerotic plaques increases the risk of ischaemic stroke, which has significant global morbidity and mortality rates. However, the specific characteristics of immune cells with dysregulated function and proven biomarkers for the diagnosis of atherosclerotic plaque progression remain poorly characterised. Our study elucidated the role of immune cells and explored diagnostic biomarkers in advanced plaque progression using single-cell RNA sequencing and high-dimensional weighted gene co-expression network analysis. We identified a subcluster of monocytes with significantly increased infiltration in the advanced plaques. Based on the monocyte signature and machine-learning approaches, we accurately distinguished advanced plaques from early plaques, with an area under the curve (AUC) of 0.899 in independent external testing. Using microenvironment cell populations (MCP) counter and non-negative matrix factorisation, we determined the association between monocyte signatures and immune cell infiltration as well as the heterogeneity of the patient. Finally, we constructed a convolutional neural network deep learning model based on gene-immune correlation, which achieved an AUC of 0.933, a sensitivity of 92.3%, and a specificity of 87.5% in independent external testing for diagnosing advanced plaques. Our findings on unique subpopulations of monocytes that contribute to carotid plaque progression are crucial for the development of diagnostic tools for clinical diseases.

Development and validation of an interpretable machine learning model for predicting the risk of distant metastasis in papillary thyroid cancer: a multicenter study

The survival rate of patients with distant metastasis (DM) of papillary thyroid carcinoma (PTC) is significantly reduced. It is of great significance to find an effective method for early prediction of the risk of DM for formulating individualized diagnosis and treatment plans and improving prognosis. Previous studies have significant limitations, and it is still necessary to develop new models for predicting the risk of DM of PTC. We aimed to develop and validate interpretable machine learning (ML) models for early prediction of DM in patients with PTC using a multicenter cohort. We collected data on patients with PTC who were admitted between June 2013 and May 2023. Data from 1430 patients at Yunnan Cancer Hospital (YCH) served as the training and internal validation set, while data from 434 patients at the First Affiliated Hospital of Kunming Medical University (KMU 1st AH) was used as the external test set. Nine ML methods such as random forest (RF) were used to construct the model. Model prediction performance was compared using evaluation indicators such as the area under the receiver operating characteristic curve (AUC). The SHapley Additive exPlanation (SHAP) method was used to rank the feature importance and explain the final model. Among the nine ML models, the RF model performed the best. The RF model accurately predicted the risk of DM in patients with PTC in both the internal validation of the training set [AUC: 0.913, 95% confidence interval (CI) (0.9075-0.9185)] and the external test set [AUC: 0.8996, 95% CI (0.8483-0.9509)]. The calibration curve showed high agreement between the predicted and observed risks. In the sensitivity analysis focusing on DM sites of PTC, the RF model exhibited outstanding performance in predicting "lung-only metastasis" showing high AUC, specificity, sensitivity, F1 score, and a low Brier score. SHAP analysis identified variables that contributed to the model predictions. An online calculator based on the RF model was developed and made available for clinicians at https://predictingdistantmetastasis.shinyapps.io/shiny1/. 11 variables were included in the final RF model: age of the patient with PTC, whether the tumor size is>2cm, whether the tumor size is≤1cm, lymphocyte (LYM) count, monocyte (MONO) count, monocyte/lymphocyte ratio (MLR), thyroglobulin (TG) level, thyroid peroxidase antibody (TPOAb) level, whether the T stage is T1/2, whether the T stage is T3/4, and whether the N stage is N0. On the basis of large-sample and multicenter data, we developed and validated an explainable ML model for predicting the risk of DM in patients with PTC. The model helps clinicians to identify high-risk patients early and provides a basis for individualized patient treatment plans. This work was supported by the National Natural Science Foundation of China (No. 81960426, 82360345 and 82001986), the Outstanding Youth Science Foundation of Yunnan Basic Research Project (No. 202401AY070001-316), Yunnan Province Applied and Basic Research Foundation (No. 202401AT070008), and Ten Thousand Talent Plans for Young Top-notch Talents of Yunnan Province.

Using Machine Learning Models to Predict Pathologic Complete Response to Neoadjuvant Chemotherapy in Breast Cancer.

Neoadjuvant chemotherapy (NAC) is increasingly used in breast cancer. Predictive modeling is useful in predicting pathologic complete response (pCR) to NAC. We test machine learning (ML) models to predict pCR in breast cancer and explore methods of handling missing data. Four hundred and ninety-nine patients with breast cancer treated with NAC in two centers in Singapore (National Cancer Centre Singapore [NCCS] and KK Hospital) between January 2014 and December 2017 were included. Eleven clinical features were used to train five different ML models. Listwise deletion and imputation were evaluated on handling missing data. Model performance was evaluated by AUC and calibration (Brier score). Feature importance from the best performing model in the external testing data set was calculated using Shapley additive explanations. Seventy-two (24.6%), 18 (24.7%), and 31 (24.8%) patients attained pCR in NCCS training, NCCS testing, and KK Women's and Children's Hospital (KKH) testing data sets, respectively. The random forest (RF) base and imputed models have the highest AUCs in the KKH cohort of 0.794 (95% CI, 0.709 to 0.873) and 0.795 (95% CI, 0.706 to 0.871), respectively, and were the best calibrated with the lowest Brier score. No statistically significant difference was noted between AUCs of the base and imputed models in all data sets. The imputed model had a larger positive predictive value (PPV; 98.2% v 95.1%) and negative predictive value (NPV; 96.7% v 90.0%) than the base model in the KKH data set. Estrogen receptor intensity, human epidermal growth factor 2 intensity, and age at diagnosis were the three most important predictors. ML, particularly RF, demonstrates reasonable accuracy in pCR prediction after NAC. Imputing missing fields in the data can improve the PPV and NPV of the pCR prediction model.

Precise Image-level Localization of Intracranial Hemorrhage on Head CT Scans with Deep Learning Models Trained on Study-level Labels.

Purpose To develop a highly generalizable weakly supervised model to automatically detect and localize image-level intracranial hemorrhage (ICH) by using study-level labels. Materials and Methods In this retrospective study, the proposed model was pretrained on the image-level Radiological Society of North America dataset and fine-tuned on a local dataset by using attention-based bidirectional long short-term memory networks. This local training dataset included 10 699 noncontrast head CT scans in 7469 patients, with ICH study-level labels extracted from radiology reports. Model performance was compared with that of two senior neuroradiologists on 100 random test scans using the McNemar test, and its generalizability was evaluated on an external independent dataset. Results The model achieved a positive predictive value (PPV) of 85.7% (95% CI: 84.0, 87.4) and an area under the receiver operating characteristic curve of 0.96 (95% CI: 0.96, 0.97) on the held-out local test set (n = 7243, 3721 female) and 89.3% (95% CI: 87.8, 90.7) and 0.96 (95% CI: 0.96, 0.97), respectively, on the external test set (n = 491, 178 female). For 100 randomly selected samples, the model achieved performance on par with two neuroradiologists, but with a significantly faster (P < .05) diagnostic time of 5.04 seconds per scan (vs 86 seconds and 22.2 seconds for the two neuroradiologists, respectively). The model's attention weights and heatmaps visually aligned with neuroradiologists' interpretations. Conclusion The proposed model demonstrated high generalizability and high PPVs, offering a valuable tool for expedited ICH detection and prioritization while reducing false-positive interruptions in radiologists' workflows. Keywords: Computer-Aided Diagnosis (CAD), Brain/Brain Stem, Hemorrhage, Convolutional Neural Network (CNN), Transfer Learning Supplemental material is available for this article. © RSNA, 2024 See also the commentary by Akinci D'Antonoli and Rudie in this issue.

In silico modeling and experimental validation of 2-oxoimidazolidin-4-sulfonamides as low-toxicity fungicides against Phytophthora infestans

Phytophthora infestans control is a long-standing problem that has caused ongoing difficulties and brought limited success for over a century. Traditional methods, such as fungicides, have drawbacks including high cost, restrictions on organic farming, potential risks to the environment and human health, and the development of resistant strains. In this study, we employed cutting-edge computer-based techniques, including Quantitative Structure-Activity Relationship (QSAR) modeling and molecular docking simulations, to uncover new fungicidal compounds and gain insights into their specific mechanisms of action against P. infestans. QSAR modeling on the number of compounds tested as P. infestans inhibitors was performed using an interactive OCHEM web platform. The predictive ability of the developed classification models had a balanced accuracy (BA) of 77–85 % for the training set and BA = 89–93 % for the validation external test set. During the in vitro testing against P. infestans, thirteen synthesized 2-oxoimidazolidine-4-sulfonamides demonstrated inhibition rates, ranging from 23.6 % to 87.4 %. The fungicidal potential of six of these fungicides ranged from 79.3 % to 87.4 %, which is comparable to the activity of known fungicides. Acute toxicity results using the well-known aquatic marker Daphnia magna showed that the most active sulfonamides 3d, 3f, 3h, 3j, 3k, and 3l, with LC50 values ranging from 13.7 to 52.9 mg/L, are low-toxicity compounds. The molecular docking results demonstrated a potential mechanism of the antifungal action of the studied 2-oxoimidazolidin-4-sulfonamide derivatives via the inhibition of fungal CYP51, a sterol biosynthesis enzyme.

Prediction of femoral head collapse in osteonecrosis using deep learning segmentation and radiomics texture analysis of MRI

BackgroundFemoral head collapse is a critical pathological change and is regarded as turning point in disease progression in osteonecrosis of the femoral head (ONFH). In this study, we aim to build an automatic femoral head collapse prediction pipeline for ONFH based on magnetic resonance imaging (MRI) radiomics.MethodsIn the segmentation model development dataset, T1-weighted MRI of 222 hips from two hospitals were retrospectively collected and randomly split into training (n = 190) and test (n = 32) sets. In the prognosis prediction model development dataset, 206 hips were also retrospectively collected from two hospitals and divided into training set (n = 155) and external test set (n = 51) according to data source. A deep learning model for automatic lesion segmentation was trained with nnU-Net, from which three-dimensional regions of interest were segmented and a total of 107 radiomics features were extracted. After intra-class correlation coefficients screening, feature correlation coefficient screening and Least Absolute Shrinkage and Selection Operator regression feature selection, a machine learning model for ONFH prognosis prediction was trained with Logistic Regression (LR) and Light Gradient Boosting Machine (LightGBM) algorithm.ResultsThe segmentation model achieved an average dice similarity coefficient of 0.848 and an average 95% Hausdorff distance of 3.794 in the test set, compared to the manual segmentation results. After feature selection, nine radiomics features were included in the prognosis prediction model. External test showed that the LightGBM model exhibited acceptable predictive performance. The area under the curve (AUC) of the prediction model was 0.851 (95% CI: 0.7268–0.9752), with an accuracy of 0.765, sensitivity of 0.833, and specificity of 0.727. Decision curve analysis showed that the LightGBM model exhibited favorable clinical utility.ConclusionThis study presents an automated pipeline for predicting femoral head collapse in ONFH with acceptable performance. Further research is necessary to determine the clinical applicability of this radiomics-based approach and to assess its potential to assist in treatment decision-making for ONFH.

Deep learning model for diagnosis of thyroid nodules with size less than 1 cm: A multicenter, retrospective study

ObjectiveTo develop a ultrasound images based dual-channel deep learning model to achieve accurate early diagnosis of thyroid nodules less than 1 cm. MethodsA dual-channel deep learning model called thyroid nodule transformer network (TNT-Net) was proposed. The model has two input channels for transverse and longitudinal ultrasound images of thyroid nodules, respectively. A total of 9649 nodules from 8455 patients across five hospitals were retrospectively collected. The data were divided into a training set (8453 nodules, 7369 patients), an internal test set (565 nodules, 512 patients), and an external test set (631 nodules, 574 patients). ResultsTNT-Net achieved an area under the curve (AUC) of 0.953 (95 % confidence interval (CI): 0.934, 0.969) on the internal test set and 0.941 (95 % CI: 0.921, 0.957) on the external test set, significantly outperforming traditional deep convolutional neural network models and single-channel swin transformer model, whose AUCs ranged from 0.800 (95 % CI: 0.759, 0.837) to 0.856 (95 % CI: 0.819, 0.881). Furthermore, feature heatmap visualization showed that TNT-Net could extract richer and more energetic malignant nodule patterns. ConclusionThe proposed TNT-Net model significantly improved the recognition capability for thyroid nodules with size less than 1 cm. This model has the potential to reduce overdiagnosis and overtreatment of such nodules, providing essential support for precise management of thyroid nodules while complementing fine-needle aspiration biopsy.

A New Intraocular Lens Power Formula Integrating an Artificial Intelligence-Powered Estimation for Effective Lens Position Based on Chinese Eyes.

To develop and evaluate a new intraocular lens (IOL) formula based on Chinese eyes. A training dataset of 709 eyes undergoing uneventful cataract surgery was used to train the algorithm for effective lens position estimation. The algorithm was then integrated with Gaussian optics to develop the new IOL formula (Jin-AI). From the same center, 177 eyes served as an internal test dataset. An independent dataset of 557 eyes served as an external test dataset. The standard deviation (SD) of prediction errors was compared among the Jin-AI formula, traditional third-generation formulas (SRK/T, Holladay 1, Hoffer Q), and newer generation formulas (Kane, Barrett Universal II [BUII], Hill-radial basis function [RBF] 3.0, and PEARL-DGS). In the internal test dataset, the Jin-AI formula showed the lowest SD (0.25 D), followed by the BUII (0.31 D), Kane (0.33 D), and PEARL-DGS (0.33 D) formulas. In the external test dataset, the Jin-AI, Kane, and PEARL-DGS formulas had the lowest SD (0.38 D), followed by the BUII (0.39 D), Hill-RBF 3.0 (0.39 D), SRK/T (0.45 D), Holladay 1 (0.48 D), and Hoffer Q (0.48 D) formulas. The SD of the Jin-AI formula was significantly lower than the third-generation formulas and comparable to the four newer generation formulas. Predictive accuracy of the Jin-AI formula was similar to the newer generation formulas across all axial length, keratometry, and anterior chamber depth ranges. The new formula has exhibited good performance in predicting postoperative refractions. Its overall predictive accuracy was better than the third-generation formulas and comparable to the newer generation ones. The Jin-AI formula could be a reliable alternative for IOL power calculation in Chinese.

Automatic tumor segmentation and lymph node metastasis prediction in papillary thyroid carcinoma using ultrasound keyframes.

Accurate preoperative prediction of cervical lymph node metastasis (LNM) for papillary thyroid carcinoma (PTC) patients is essential for disease staging and individualized treatment planning, which can improve prognosis and facilitate better management. To establish a fully automated deep learning-enabled model (FADLM) for automated tumor segmentation and cervical LNM prediction in PTC using ultrasound (US) video keyframes. The bicentral study retrospective enrolled 518 PTC patients, who were then randomly divided into the training (Hospital 1, n=340), internal test (Hospital 1, n=83), and external test cohorts (Hospital 2, n=95). The FADLM integrated mask region-based convolutional neural network (Mask R-CNN) for automatic thyroid primary tumor segmentation and ResNet34 with Bayes strategy for cervical LNM diagnosis. A radiomics model (RM) using the same automated segmentation method, a traditional radiomics model (TRM) using manual segmentation, and a clinical-semantic model (CSM) were developed for comparison. The dice similarity coefficient (DSC) was used to evaluate segmentation performance. The prediction performance of the models was validated in terms of discrimination and clinical utility with the area under the receiver operator characteristic curve (AUC), heatmap analysis, and decision curve analysis (DCA). The comparison of the predictive performance among different models was conducted by DeLong test. The performances of two radiologists compared with FADLM and the diagnostic augmentation with FADLM's assistance were analyzed in terms of accuracy, sensitivity and specificity using McNemar's x2 test. The p-value less than 0.05 was defined as a statistically significant difference. The Benjamini-Hochberg procedure was applied for multiple comparisons to deal with Type I error. The FADLM yielded promising segmentation results in training (DSC: 0.88±0.23), internal test (DSC: 0.88±0.23), and external test cohorts (DSC: 0.85±0.24). The AUCs of FADLM for cervical LNM prediction were 0.78 (95% CI: 0.73, 0.83), 0.83 (95% CI: 0.74, 0.92), and 0.83 (95% CI: 0.75, 0.92), respectively. It all significantly outperformed the RM (AUCs: 0.78vs. 0.72; 0.83vs. 0.65; 0.83vs. 0.68, all adjusted p-values<0.05) and CSM (AUCs: 0.78vs. 0.71; 0.83vs. 0.62; 0.83vs. 0.68, all adjusted p-values<0.05) across the three cohorts. The RM offered similar performance to that of TRM (AUCs: 0.61vs. 0.63, adjusted p-value =0.60) while significantly reducing the segmentation time (3.3±3.8 vs. 14.1±4.2 s, p-value <0.001). Under the assistance of FADLM, the accuracies of junior and senior radiologists were improved by 18% and 15% (all adjusted p-values<0.05) and the sensitivities by 25% and 21% (all adjusted p-values<0.05) in the external test cohort. The FADLM with elaborately designed automated strategy using US video keyframes holds good potential to provide an efficient and consistent prediction of cervical LNM in PTC. The FADLM displays superior performance to RM, CSM, and radiologists with promising efficacy.

Using Joint Longitudinal and Time-to-Event Models to Improve the Parameterization of Chronic Disease Microsimulation Models: an Application to Cardiovascular Disease.

Chronic disease microsimulation models often simulate disease incidence as a function of risk factors that evolve over time (e.g., blood pressure increasing with age) in order to facilitate decision analyses of different disease screening and prevention strategies. Existing models typically rely on incidence rates estimated with standard survival analysis techniques (e.g., proportional hazards from baseline data) that are not designed to be continually updated each model cycle. We introduce the use of joint longitudinal and time-to-event to parameterize microsimulations to avoid potential issues from using these existing methods. These joint models include random effects regressions to estimate the risk factor trajectories and a survival model to predict disease risk based on those estimated trajectories. In a case study on cardiovascular disease (CVD), we compare the validity of microsimulation models parameterized with this joint model approach to those parameterized with the standard approaches. A CVD microsimulation model was constructed that modeled the trajectory of seven CVD risk factors/predictors as a function of age (smoking, diabetes, systolic blood pressure, antihypertensive medication use, total cholesterol, HDL, and statin use) and predicted yearly CVD incidence as a function of these predictors, plus age, sex, and race. We parameterized the model using data from the Atherosclerosis Risk in Communities study (ARIC). The risk of CVD in the microsimulation was parameterized with three approaches: (1) joint longitudinal and time-to-event model, (2) proportional hazards model estimated using baseline data, and (3) proportional hazards model estimated using time-varying data. We accounted for non-CVD mortality across all the parameterization approaches. We simulated risk factor trajectories and CVD incidence from age 70y to 85y for an external test set comprised of individuals from the Multi-Ethnic Study of Atherosclerosis (MESA). We compared the simulated to observed incidence using both average survival curves and the E50 and E90 calibration metrics (the median and 90th percentile absolute difference between observed and predicted incidence) to measure the validity of each parameterization approach. The average CVD survival curve estimated by the microsimulation model parameterized with the joint model approach matched the observed curve from the test set relatively closely. The other parameterization methods generally performed worse, especially the proportional hazards model estimated using baseline data. Similar results were observed for the calibration metrics, with the joint model performing particularly well on the E90 metric compared to the other models. Using a joint longitudinal and time-to-event model to parameterize a CVD simulation model produced incidence predictions that more accurately reflected observed data than a model parameterized with standard approaches. This parameterization approach could lead to more reliable microsimulation models, especially for models that evaluate policies which depend on tracking dynamic risk factors over time. Beyond this single case study, more work is needed to identify the specific circumstances where the joint model approach will outperform existing methods.

Quantitative texture analysis using machine learning for predicting interpretable pulmonary perfusion from non-contrast computed tomography in pulmonary embolism patients

BackgroundPulmonary embolism (PE) is life-threatening and requires timely and accurate diagnosis, yet current imaging methods, like computed tomography pulmonary angiography, present limitations, particularly for patients with contraindications to iodinated contrast agents. We aimed to develop a quantitative texture analysis pipeline using machine learning (ML) based on non-contrast thoracic computed tomography (CT) scans to discover intensity and textural features correlated with regional lung perfusion (Q) physiology and pathology and synthesize voxel-wise Q surrogates to assist in PE diagnosis.MethodsWe retrospectively collected 99mTc-labeled macroaggregated albumin Q-SPECT/CT scans from patients suspected of PE, including an internal dataset of 76 patients (64 for training, 12 for testing) and an external testing dataset of 49 patients. Quantitative CT features were extracted from segmented lung subregions and underwent a two-stage feature selection pipeline. The prior-knowledge-driven preselection stage screened for robust and non-redundant perfusion-correlated features, while the data-driven selection stage further filtered features by fitting ML models for classification. The final classification model, trained with the highest-performing PE-associated feature combination, was evaluated in the testing cohorts based on the Area Under the Curve (AUC) for subregion-level predictability. The voxel-wise Q surrogate was then synthesized using the final selected feature maps (FMs) and model score maps (MSMs) to investigate spatial distributions. The Spearman correlation coefficient (SCC) and Dice similarity coefficient (DSC) were used to assess the spatial consistency between FMs or MSMs and Q-SPECT scans.ResultsThe optimal model performance achieved an AUC of 0.863 during internal testing and 0.828 on the external testing cohort. The model identified a combination containing 14 intensity and textural features that were non-redundant, robust, and capable of distinguishing between high- and low-functional lung regions. Spatial consistency assessment in the internal testing cohort showed moderate-to-high agreement between MSMs and reference Q-SPECT scans, with median SCC of 0.66, median DSCs of 0.86 and 0.64 for high- and low-functional regions, respectively.ConclusionsThis study validated the feasibility of using quantitative texture analysis and a data-driven ML pipeline to generate voxel-wise lung perfusion surrogates, providing a radiation-free, widely accessible alternative to functional lung imaging in managing pulmonary vascular diseases.Clinical trial numberNot applicable.

Machine learning approach for detection of valvular heart disease through digital cardiac auscultation: results from multi-center prospective cross-sectional study

Abstract Background Cardiac auscultation, an essential bedside diagnostic tool, often identifies significant valvular heart disease through murmurs. However, its accuracy is contingent upon the clinician's expertise, leading to a growing dependence on expensive imaging techniques. This study aimed to develop and validate a machine learning model for detecting valvular heart disease from cardiac auscultation. Methods The digital cardiac auscultation and associated echocardiograms were prospectively collected from three sites between May 1, 2022, and Dec 31, 2022. We trained, validated, and externally tested a machine learning based model to classify aortic stenosis, mitral stenosis, aortic regurgitation, mitral regurgitation, tricuspid regurgitation, oveall valvular heart disease from cardiac auscultation, age, sex, and comorbidity. We condcted training, and internal testing using datasets for eXtreme Gradient Boosting (XGBoost) machine learning model from hospital one and hospital two. The dataset from hospital three was employed for external testing. We evaluated the area under the receiver operating characteristic curve (AUROC), area under the precision recall curve (AUPRC), accuracy, sensitivity, specificity. Result We selected the cardiac auscultation and echocardiograms of a total of 4,321 patients, excluding those without valvular heart disease information, congenital heart disease, a history of valvular operation, and cardiac auscultation could not be interpreted. The diagnostic performance of heart murmurs for valvular heart disease across the three datasets generally demonstrated high sensitivity and PPV. Specifically, the sensitivity for aortic stenosis was 89.1%, mitral stenosis 88.8%, aortic regurgitation 88.5%, mitral regurgitation 89.8%, tricuspid regurgitation 90.1%, and for any valvular heart disease, it was 95.0%. However, the NPV was consistently lower, with values of 16.4%, 14.2%, 12.2%, 23.1%, 26.2%, and 65.3% for the respective conditions. In XGBoost machine learning model, the AUROC values for internal testing (N=1,403), along with their 95% confidence intervals, were as follows: aortic stenosis (0.93, 0.88–0.97), mitral stenosis (0.91, 0.81–0.97), aortic regurgitation (0.84, 0.79–0.90), mitral regurgitation (0.94, 0.90–0.98), tricuspid regurgitation (0.81, 0.75–0.87), and overall valvular heart disease (0.86, 0.82–0.89). For external testing (N=814), the AUROC values were: aortic stenosis (0.90, 0.83–0.96), mitral stenosis (0.88, 0.79–0.95), aortic regurgitation (0.94, 0.93–0.96), mitral regurgitation (0.93, 0.92–0.95), tricuspid regurgitation (0.89, 0.87–0.91), and overall valvular heart disease (0.92, 0.90–0.94). Conclusion The machine learning based model can accurately classify valvular heart diseases using information from digital cardiac auscultation.Figure 1.Study FlowFigure 2.AUC curve

Computational staining of CD3/CD20 positive lymphocytes in human tissues with experimental confirmation in a genetically engineered mouse model.

Immune dysregulation plays a major role in cancer progression. The quantification of lymphocytic spatial inflammation may enable spatial system biology, improve understanding of therapeutic resistance, and contribute to prognostic imaging biomarkers. In this paper, we propose a knowledge-guided deep learning framework to measure the lymphocytic spatial architecture on human H&E tissue, where the fidelity of training labels is maximized through single-cell resolution image registration of H&E to IHC. We demonstrate that such an approach enables pixel-perfect ground-truth labeling of lymphocytes on H&E as measured by IHC. We then experimentally validate our technique in a genetically engineered, immune-compromised Rag2 mouse model, where Rag2 knockout mice lacking mature lymphocytes are used as a negative experimental control. Such experimental validation moves beyond the classical statistical testing of deep learning models and demonstrates feasibility of more rigorous validation strategies that integrate computational science and basic science. Using our developed approach, we automatically annotated more than 111,000 human nuclei (45,611 CD3/CD20 positive lymphocytes) on H&E images to develop our model, which achieved an AUC of 0.78 and 0.71 on internal hold-out testing data and external testing on an independent dataset, respectively. As a measure of the global spatial architecture of the lymphocytic microenvironment, the average structural similarity between predicted lymphocytic density maps and ground truth lymphocytic density maps was 0.86 ± 0.06 on testing data. On experimental mouse model validation, we measured a lymphocytic density of 96.5 ± %1% in a Rag2+/- control mouse, compared to an average of 16.2 ± %5% in Rag2-/- immune knockout mice (p<0.0001, ANOVA-test). These results demonstrate that CD3/CD20 positive lymphocytes can be accurately detected and characterized on H&E by deep learning and generalized across species. Collectively, these data suggest that our understanding of complex biological systems may benefit from computationally-derived spatial analysis, as well as integration of computational science and basic science.

SDS-Net: A Synchronized Dual-Stage Network for Predicting Patients Within 4.5-h Thrombolytic Treatment Window Using MRI.

Timely and precise identification of acute ischemic stroke (AIS) within 4.5h is imperative for effective treatment decision-making. This study aims to construct a novel network that utilizes limited datasets to recognize AIS patients within this critical window. We conducted a retrospective analysis of 265 AIS patients who underwent both fluid attenuation inversion recovery (FLAIR) and diffusion-weighted imaging (DWI) within 24h of symptom onset. Patients were categorized based on the time since stroke onset (TSS) into two groups: TSS ≤ 4.5 h and TSS > 4.5 h. The TSS was calculated as the time from stroke onset to MRI completion. We proposed a synchronized dual-stage network (SDS-Net) and a sequential dual-stage network (Dual-stage Net), which were comprised of infarct voxel identification and TSS classification stages. The models were trained on 181 patients and validated on an independent external cohort of 84 patients using metrics of area under the curve (AUC), sensitivity, specificity, and accuracy. A DeLong test was used to statistically compare the performance of the two models. SDS-Net achieved an accuracy of 0.844 with an AUC of 0.914 in the validation dataset, outperforming the Dual-stage Net, which had an accuracy of 0.822 and an AUC of 0.846. In the external test dataset, SDS-Net further demonstrated superior performance with an accuracy of 0.800 and an AUC of 0.879, compared to the accuracy of 0.694 and AUC of 0.744 of Dual-stage Net (P = 0.049). SDS-Net is a robust and reliable tool for identifying AIS patients within a 4.5-h treatment window using MRI. This model can assist clinicians in making timely treatment decisions, potentially improving patient outcomes.

Deep learning enables the fully automated quantification of high-resolution cardiac CT images: human in loop for model development

Abstract Introduction Cardiac computed tomography (CCT) images provide vast spatial and temporal information. Segmentation of the structure of interest is necessary to provide quantitative information. However, manual segmentation of cardiac substructures is highly time-consuming. Objective This study aims to develop a robust deep-learning segmentation model for cardiac substructure segmentation in 3D CCT images. Method We initially developed a 3D U-Net architecture, incorporating extensive augmentation, for the segmentation of the Left Ventricle Myocardium (LVM), the blood cavities of the Left and Right Ventricles (LV, RV), and the Left and Right Atria (LA, RA). This model utilized CT images from 20 patients who had available manual segmentation. Subsequently, we applied this initial model to datasets from two centers, including different generations of scanners. We then visually inspected the segmentation, selected the highest-quality segmentations, and modify it if necessary, and excluded the low-quality data. We continued fine-tuning the model with the new dataset until we reached a count of 456 patients. As there were multiple images from one patient during this model's development, we generated datasets to develop a new model using the entire new dataset to avoid any data leakage in the reported metrics. The model was developed using data from Center 1 (5-fold cross-validation) and then evaluated on an external test set from Center 2 (126 patients) and also 4D images from Center 1. Various metrics were calculated to assess the model's performance. Results In the internal test set, utilizing 5-fold CV, the model demonstrated high segmentation accuracy, achieving an overall Dice coefficient of 0.98, with individual scores of 0.96 for LVM, 0.98 for LV, 0.98 for RV, 0.99 for RA, and 0.98 for LA. We assessed the automatic quantification capability of the segmentation model from 256 patients, resulting in an R2 value greater than 0.92 for both LV end-diastolic and systolic volumes. In contrast, the model maintained robust performance on the external dataset, achieving an overall Dice coefficient of 0.96, with individual scores of 0.95 for LVM, 0.98 for LV, 0.97 for RV, 0.98 for RA, and 0.94 for LA. Conclusion We provided a large cohort of ground truth segmentations for different cardiac CT substructures using a human-in-the-loop strategy. Subsequently, we developed a deep learning model for highly accurate automated segmentation of these substructures. This model allows for fast and fully automatic segmentation and quantification of 3D and 4D cardiac CT images, from which various quantitative metrics can be extracted from different cardiac substructures.

An interpretable deep learning model for detecting BRCA pathogenic variants of breast cancer from hematoxylin and eosin-stained pathological images.

Determining the status of breast cancer susceptibility genes (BRCA) is crucial for guiding breast cancer treatment. Nevertheless, the need for BRCA genetic testing among breast cancer patients remains unmet due to high costs and limited resources. This study aimed to develop a Bi-directional Self-Attention Multiple Instance Learning (BiAMIL) algorithm to detect BRCA status from hematoxylin and eosin (H&E) pathological images. A total of 319 histopathological slides from 254 breast cancer patients were included, comprising two dependent cohorts. Following image pre-processing, 633,484 tumor tiles from the training dataset were employed to train the self-developed deep-learning model. The performance of the network was evaluated in the internal and external test sets. BiAMIL achieved AUC values of 0.819 (95% CI [0.673-0.965]) in the internal test set, and 0.817 (95% CI [0.712-0.923]) in the external test set. To explore the relationship between BRCA status and interpretable morphological features in pathological images, we utilized Class Activation Mapping (CAM) technique and cluster analysis to investigate the connections between BRCA gene mutation status and tissue and cell features. Significantly, we observed that tumor-infiltrating lymphocytes and the morphological characteristics of tumor cells appeared to be potential features associated with BRCA status. An interpretable deep neural network model based on the attention mechanism was developed to predict the BRCA status in breast cancer. Keywords: Breast cancer, BRCA, deep learning, self-attention, interpretability.

Prediction, prognosis and monitoring of neurodegeneration at biobank-scale via machine learning and imaging.

Alzheimer's disease and related dementias (ADRD) and Parkinson's disease (PD) are the most common neurodegenerative conditions. These central nervous system disorders impact both the structure and function of the brain and may lead to imaging changes that precede symptoms. Patients with ADRD or PD have long asymptomatic phases that exhibit significant heterogeneity. Hence, quantitative measures that can provide early disease indicators are necessary to improve patient stratification, clinical care, and clinical trial design. This work uses machine learning techniques to derive such a quantitative marker from T1-weighted (T1w) brain Magnetic resonance imaging (MRI). In this retrospective study, we developed machine learning (ML) based disease-specific scores based on T1w brain MRI utilizing Parkinson's Disease Progression Marker Initiative (PPMI) and Alzheimer's Disease Neuroimaging Initiative (ADNI) cohorts. We evaluated the potential of ML-based scores for early diagnosis, prognosis, and monitoring of ADRD and PD in an independent large-scale population-based longitudinal cohort, UK Biobank. 1,826 dementia images from 731 participants, 3,161 healthy control images from 925 participants from the ADNI cohort, 684 PD images from 319 participants, and 232 healthy control images from 145 participants from the PPMI cohort were used to train machine learning models. The classification performance is 0.94 [95% CI: 0.93-0.96] area under the ROC Curve (AUC) for ADRD detection and 0.63 [95% CI: 0.57-0.71] for PD detection using 790 extracted structural brain features. The most predictive regions include the hippocampus and temporal brain regions in ADRD and the substantia nigra in PD. The normalized ML model's probabilistic output (ADRD and PD imaging scores) was evaluated on 42,835 participants with imaging data from the UK Biobank. There are 66 cases for ADRD and 40 PD cases whose T1 brain MRI is available during pre-diagnostic phases. For diagnosis occurrence events within 5 years, the integrated survival model achieves a time-dependent AUC of 0.86 [95% CI: 0.80-0.92] for dementia and 0.89 [95% CI: 0.85-0.94] for PD. ADRD imaging score is strongly associated with dementia-free survival (hazard ratio (HR) 1.76 [95% CI: 1.50-2.05] per S.D. of imaging score), and PD imaging score shows association with PD-free survival (hazard ratio 2.33 [95% CI: 1.55-3.50]) in our integrated model. HR and prevalence increased stepwise over imaging score quartiles for PD, demonstrating heterogeneity. As a proxy for diagnosis, we validated AD/PD polygenic risk scores of 42,835 subjects against the imaging scores, showing a highly significant association after adjusting for covariates. In both the PPMI and ADNI cohorts, the scores are associated with clinical assessments, including the Mini-Mental State Examination (MMSE), Alzheimer's Disease Assessment Scale-cognitive subscale (ADAS-Cog), and pathological markers, which include amyloid and tau. Finally, imaging scores are associated with polygenic risk scores for multiple diseases. Our results suggest that we can use imaging scores to assess the genetic architecture of such disorders in the future. Our study demonstrates the use of quantitative markers generated using machine learning techniques for ADRD and PD. We show that disease probability scores obtained from brain structural features are useful for early detection, prognosis prediction, and monitoring disease progression. To facilitate community engagement and external tests of model utility, an interactive app to explore summary level data from this study and dive into external data can be found here https://ndds-brainimaging-ml.streamlit.app. As far as we know, this is the first publicly available cloud-based MRI prediction application. US National Institute on Aging, and US National Institutes of Health.

Automatic identification of coronary stent in coronary calcium scoring CT using deep learning

Coronary artery calcium (CAC) scoring CT is a useful tool for screening coronary artery disease and for cardiovascular risk stratification. However, its efficacy in patients with coronary stents, who had pre-existing coronary artery disease, remains uncertain. Historically, CAC CT scans of these patients have been manually excluded from the CAC scoring process, even though most of the CAC scoring process is now fully automated. Therefore, we hypothesized that automating the filtering of patients with coronary stents using artificial intelligence could streamline the entire CAC workflow, eliminating the need for manual intervention. Consequently, we aimed to develop and evaluate a deep learning-based coronary stent filtering algorithm (StentFilter) in CAC scoring CT scans using a multicenter CAC dataset. We developed StentFilter comprising two main processes: stent identification and false-positive reduction. Development utilized 108 non-enhanced echocardiography-gated CAC scans (including 74 with manually labeled stents), and for false positive reduction, 2063 CAC scans with significant coronary calcium (average Agatston score: 523.8) but no stents were utilized. StentFilter’s performance was evaluated on two independent internal test sets (Asan cohort- and 2; n = 355 and 396; one without coronary stents) and two external test sets from different institutions (n = 105 and 62), each with manually labeled stents. We calculated the per-patient sensitivity, specificity, and false-positive rate of StentFilter. StentFilter demonstrated a high overall per-patient sensitivity of 98.8% (511/517 cases with stents) and a false-positive rate of 0.022 (20/918). Notably, the false-positive ratio was significantly lower in the dataset containing stents (Asan cohort-1; 0.008 [3/355]) compared with the dataset without stents (Asan cohort-2; 0.043 [17/396], p = 0.008). All false-positive identifications were attributed to dense coronary calcifications, with no false positives identified in extracoronary locations. The automated StentFilter accurately distinguished coronary stents from pre-existing coronary calcifications. This approach holds potential as a filter within a fully automated CAC scoring workflow, streamlining the process efficiently.

Construction and validation of a line chart for gestational diabetes mellitus based on clinical indicators

BackgroundGestational diabetes mellitus (GDM) is a common complication of mid–to-late pregnancy. Here, we constructed a predictive model for GDM based on a combination of clinical characteristics and relevant serum markers.MethodsData from full-term singleton vaginal deliveries from January 2022 to January 2023 were retrospectively collected from the obstetrics department. The data collected were segregated and assigned to training, validation, and external test sets. Maternal demographic characteristics, living and working habits, and haematological indicators, such as liver function and lipids were collected using a questionnaire designed for the study. The “rms” package in R was used to explore GDM-associated factors through stepwise regression at P < 0.05. A predictive model was developed based on the results of multifactorial logistic regression analysis. We then evaluated the differentiation of the column-line graphical model and performed internal and external validation. To assess the accuracy of the bar graphical model, we plotted calibration and decision curves.ResultsData from 265 pregnant women were included in the training and internal validation sets, and data from 113 pregnant women were included in the external validation set. The logistic regression algorithm screened 8 indicators as predictors. A prediction model was constructed with ALT, TBA, TC, and TG levels while considering whether GDM affects appetite, the husband– wife relationship, family history, and parental relationships as predictors. The Hosmer–Lemeshow goodness-of-fit test revealed that the chi-square values for the modelling, internal validation, and external validation groups (χ2 = 5.964, 3.249, and 12.182, respectively) were all P > 0.05. The ROC curve AUCs for the three groups were 0.93 (95% CI: 0.89–0.97), 0.72 (95% CI: 0.62–0.81), and 0.68 (95% CI: 0.53–0.83), respectively.ConclusionIn this study, a GDM prediction model was constructed to achieve high performance in GDM risk prediction based on routine obstetric tests and information.