Screening and staging of chronic obstructive pulmonary disease with deep learning based on chest X-ray images and clinical parameters.

BACKGROUND: Chronic obstructive pulmonary disease (COPD) is underdiagnosed with the current gold standard measure pulmonary function test (PFT). A more sensitive and simple option for early detection and severity evaluation of COPD could benefit practitioners and patients. METHODS: In this multicenter retrospective study, frontal chest X-ray (CXR) images and related clinical information of 1055 participants were collected and processed. Different deep learning algorithms and transfer learning models were trained to classify COPD based on clinical data and CXR images from 666 subjects, and validated in internal test set based on 284 participants. External test including 105 participants was also performed to verify the generalization ability of the learning algorithms in diagnosing COPD. Meanwhile, the model was further used to evaluate disease severity of COPD by predicting different grads. RESULTS: The Ensemble model showed an AUC of 0.969 in distinguishing COPD by simultaneously extracting fusion features of clinical parameters and CXR images in internal test, better than models that used clinical parameters (AUC = 0.963) or images (AUC = 0.946) only. For the external test set, the AUC slightly declined to 0.934 in predicting COPD based on clinical parameters and CXR images. When applying the Ensemble model to determine disease severity of COPD, the AUC reached 0.894 for three-classification and 0.852 for five-classification respectively. CONCLUSION: The present study used DL algorithms to screen COPD and predict disease severity based on CXR imaging and clinical parameters. The models showed good performance and the approach might be an effective case-finding tool with low radiation dose for COPD diagnosis and staging.

Deep Learning-Based Classification of Hepatocellular Nodular Lesions on Whole-Slide Histopathologic Images.

BACKGROUND & AIMS: Hepatocellular nodular lesions (HNLs) constitute a heterogeneous group of disorders. Differential diagnosis among these lesions, especially high-grade dysplastic nodules (HGDNs) and well-differentiated hepatocellular carcinoma (WD-HCC), can be challenging, let alone biopsy specimens. We aimed to develop a deep learning system to solve these puzzles, improving the histopathologic diagnosis of HNLs (WD-HCC, HGDN, low-grade DN, focal nodular hyperplasia, hepatocellular adenoma), and background tissues (nodular cirrhosis, normal liver tissue). METHODS: The samples consisting of surgical and biopsy specimens were collected from 6 hospitals. Each specimen was reviewed by 2 to 3 subspecialists. Four deep neural networks (ResNet50, InceptionV3, Xception, and the Ensemble) were used. Their performances were evaluated by confusion matrix, receiver operating characteristic curve, classification map, and heat map. The predictive efficiency of the optimal model was further verified by comparing with that of 9 pathologists. RESULTS: We obtained 213,280 patches from 1115 whole-slide images of 738 patients. An optimal model was finally chosen based on F1 score and area under the curve value, named hepatocellular-nodular artificial intelligence model (HnAIM), with the overall 7-category area under the curve of 0.935 in the independent external validation cohort. For biopsy specimens, the agreement rate with subspecialists' majority opinion was higher for HnAIM than 9 pathologists on both patch level and whole-slide images level. CONCLUSIONS: We first developed a deep learning diagnostic model for HNLs, which performed well and contributed to enhancing the diagnosis rate of early HCC and risk stratification of patients with HNLs. Furthermore, HnAIM had significant advantages in patch-level recognition, with important diagnostic implications for fragmentary or scarce biopsy specimens.

Deep learning based on susceptibility-weighted MR sequence for detecting cerebral microbleeds and classifying cerebral small vessel disease.

BACKGROUND: Cerebral microbleeds (CMBs) serve as neuroimaging biomarkers to assess risk of intracerebral hemorrhage and diagnose cerebral small vessel disease (CSVD). Therefore, detecting CMBs can evaluate the risk of intracerebral hemorrhage and use its presence to support CSVD classification, both are conducive to optimizing CSVD management. This study aimed to develop and test a deep learning (DL) model based on susceptibility-weighted MR sequence (SWS) to detect CMBs and classify CSVD to assist neurologists in optimizing CSVD management. Patients with arteriolosclerosis (aSVD), cerebral amyloid angiopathy (CAA), and cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy (CADASIL) treated at three centers were enrolled between January 2017 and May 2022 in this retrospective study. The SWSs of patients from two centers were used as the development set, and the SWSs of patients from the remaining center were used as the external test set. The DL model contains a Mask R-CNN for detecting CMBs and a multi-instance learning (MIL) network for classifying CSVD. The metrics for model performance included intersection over union (IoU), Dice score, recall, confusion matrices, receiver operating characteristic curve (ROC) analysis, accuracy, precision, and F1-score. RESULTS: A total of 364 SWS were recruited, including 336 in the development set and 28 in the external test set. IoU for the model was 0.523 ± 0.319, Dice score 0.627 ± 0.296, and recall 0.706 ± 0.365 for CMBs detection in the external test set. For CSVD classification, the model achieved a weighted-average AUC of 0.908 (95% CI 0.895-0.921), accuracy of 0.819 (95% CI 0.768-0.870), weighted-average precision of 0.864 (95% CI 0.831-0.897), and weighted-average F1-score of 0.829 (95% CI 0.782-0.876) in the external set, outperforming the performance of the neurologist group. CONCLUSION: The DL model based on SWS can detect CMBs and classify CSVD, thereby assisting neurologists in optimizing CSVD management.

Deep learning-enabled classification of kidney allograft rejection on whole slide histopathologic images.

Background: Diagnosis of kidney transplant rejection currently relies on manual histopathological assessment, which is subjective and susceptible to inter-observer variability, leading to limited reproducibility. We aim to develop a deep learning system for automated assessment of whole-slide images (WSIs) from kidney allograft biopsies to enable detection and subtyping of rejection and to predict the prognosis of rejection. Method: We collected H&E-stained WSIs of kidney allograft biopsies at 400x magnification from January 2015 to September 2023 at two hospitals. These biopsy specimens were classified as T cell-mediated rejection, antibody-mediated rejection, and other lesions based on the consensus reached by two experienced transplant pathologists. To achieve feature extraction, feature aggregation, and global classification, we employed multi-instance learning and common convolution neural networks (CNNs). The performance of the developed models was evaluated using various metrics, including confusion matrix, receiver operating characteristic curves, the area under the curve (AUC), classification map, heat map, and pathologist-machine confrontations. Results: In total, 906 WSIs from 302 kidney allograft biopsies were included for analysis. The model based on multi-instance learning enables detection and subtyping of rejection, named renal rejection artificial intelligence model (RRAIM), with the overall 3-category AUC of 0.798 in the independent test set, which is superior to that of three transplant pathologists under nearly routine assessment conditions. Moreover, the prognosis models accurately predicted graft loss within 1 year following rejection and treatment response for rejection, achieving AUC of 0.936 and 0.756, respectively. Conclusion: We first developed deep-learning models utilizing multi-instance learning for the detection and subtyping of rejection and prediction of rejection prognosis in kidney allograft biopsies. These models performed well and may be useful in assisting the pathological diagnosis.

Assessing breast disease with deep learning model using bimodal bi-view ultrasound images and clinical information.

Breast cancer is the second leading cause of carcinoma-linked death in women. We developed a multi-modal deep-learning model (BreNet) to differentiate breast cancer from benign lesions. BreNet was constructed and trained on 10,108 images from one center and tested on 3,762 images from two centers in three steps. The diagnostic ability of BreNet was first compared with that of six radiologists; a BreNet-aided scheme was constructed to improve the diagnostic ability of the radiologists; and the diagnosis of real-world radiologists' scheme was then compared with the BreNet-aided scheme. The diagnostic performance of BreNet was superior to that of the radiologists (area under the curve [AUC]: 0.996 vs. 0.841). BreNet-aided scheme increased the pooled AUC of the radiologists from 0.841 to 0.934 for reviewing images, and from 0.892 to 0.934 in the real-world test. The use of BreNet significantly enhances the diagnostic ability of radiologists in the detection of breast cancer.

Early prognostication of critical patients with spinal cord injury: A machine learning study with 1485 cases.

STUDY DESIGN: A retrospective case-series. OBJECTIVE: The study aims to use machine-learning (ML) to predict the discharge destination of spinal cord injury (SCI) patients in the intensive care unit (ICU). SUMMARY OF BACKGROUND DATA: Prognostication following SCI is vital, especially for critical patients who need intensive care. METHODS: Clinical data of patients diagnosed with SCI were extracted from a publicly available ICU database. The firstly recorded data of the included patients were used to develop a total of 98 ML classifiers, seeking to predict discharge destination (e.g. death, further medical care, home). The micro-average area under the curve (AUC) was the main indicator to assess discrimination. The best average-AUC classifier and the best death-sensitivity classifier were integrated into an ensemble classifier. The discrimination of the ensemble classifier was compared with top death-sensitivity classifiers and top average-AUC classifiers. Additionally, prediction consistency and clinical utility were also assessed. RESULTS: A total of 1485 SCI patients were included. The ensemble classifier had a micro-average AUC of 0.851, which was only slightly inferior to the best average-AUC classifier (P=0.10) The best average-AUC classifier death sensitivity was much lower than that of the ensemble classifier. The ensemble classifier had a death sensitivity of 0.452, which was inferior to top 8 death-sensitivity classifiers, whose micro-average AUC were inferior to the ensemble classifier (P<0.05). Additionally, the ensemble classifier demonstrated a comparable Brier score and superior Net benefit in the decision curve analysis, when compared to the performance of the origin classifiers. CONCLUSIONS: The ensemble classifier shows an overall superior performance in predicting discharge destination considering discrimination ability, prediction consistency and clinical utility. This classifier system may aid in the clinical management of critical SCI patients in the early phase following injury. LEVEL OF EVIDENCE: 3.

Automated Breast Volume Scanner (ABVS)-Based Radiomic Nomogram: A Potential Tool for Reducing Unnecessary Biopsies of BI-RADS 4 Lesions.

Improving the assessment of breast imaging reporting and data system (BI-RADS) 4 lesions and reducing unnecessary biopsies are urgent clinical issues. In this prospective study, a radiomic nomogram based on the automated breast volume scanner (ABVS) was constructed to identify benign and malignant BI-RADS 4 lesions and evaluate its value in reducing unnecessary biopsies. A total of 223 histologically confirmed BI-RADS 4 lesions were enrolled and assigned to the training and validation cohorts. A radiomic score was generated from the axial, sagittal, and coronal ABVS images. Combining the radiomic score and clinical-ultrasound factors, a radiomic nomogram was developed by multivariate logistic regression analysis. The nomogram integrating the radiomic score, lesion size, and BI-RADS 4 subcategories showed good discrimination between malignant and benign BI-RADS 4 lesions in the training (AUC, 0.959) and validation (AUC, 0.925) cohorts. Moreover, 42.5% of unnecessary biopsies would be reduced by using the nomogram, but nine (4%) malignant BI-RADS 4 lesions were unfortunately missed, of which 4A (77.8%) and small-sized (<10 mm) lesions (66.7%) accounted for the majority. The ABVS radiomics nomogram may be a potential tool to reduce unnecessary biopsies of BI-RADS 4 lesions, but its ability to detect small BI-RADS 4A lesions needs to be improved.

Machine Learning-based Prediction of Prolonged Intensive Care Unit Stay for Critical Patients with Spinal Cord Injury.

STUDY DESIGN: A retrospective cohort study. OBJECTIVE: The objective of the study was to develop machine-learning (ML) classifiers for predicting prolonged intensive care unit (ICU)-stay and prolonged hospital-stay for critical patients with spinal cord injury (SCI). SUMMARY OF BACKGROUND DATA: Critical patients with SCI in ICU need more attention. SCI patients with prolonged stay in ICU usually occupy vast medical resources and hinder the rehabilitation deployment. METHODS: A total of 1599 critical patients with SCI were included in the study and labeled with prolonged stay or normal stay. All data were extracted from the eICU Collaborative Research Database and the Medical Information Mart for Intensive Care III-IV Database. The extracted data were randomly divided into training, validation and testing (6:2:2) subdatasets. A total of 91 initial ML classifiers were developed, and the top three initial classifiers with the best performance were further stacked into an ensemble classifier with logistic regressor. The area under the curve (AUC) was the main indicator to assess the prediction performance of all classifiers. The primary predicting outcome was prolonged ICU-stay, while the secondary predicting outcome was prolonged hospital-stay. RESULTS: In predicting prolonged ICU-stay, the AUC of the ensemble classifier was 0.864 ±â€Š0.021 in the three-time five-fold cross-validation and 0.802 in the independent testing. In predicting prolonged hospital-stay, the AUC of the ensemble classifier was 0.815 ±â€Š0.037 in the three-time five-fold cross-validation and 0.799 in the independent testing. Decision curve analysis showed the merits of the ensemble classifiers, as the curves of the top three initial classifiers varied a lot in either predicting prolonged ICU-stay or discriminating prolonged hospital-stay. CONCLUSION: The ensemble classifiers successfully predict the prolonged ICU-stay and the prolonged hospital-stay, which showed a high potential of assisting physicians in managing SCI patients in ICU and make full use of medical resources.Level of Evidence: 3.

Deep Learning Networks Accurately Detect ST-Segment Elevation Myocardial Infarction and Culprit Vessel.

Early diagnosis of acute ST-segment elevation myocardial infarction (STEMI) and early determination of the culprit vessel are associated with a better clinical outcome. We developed three deep learning (DL) models for detecting STEMIs and culprit vessels based on 12-lead electrocardiography (ECG) and compared them with conclusions of experienced doctors, including cardiologists, emergency physicians, and internists. After screening the coronary angiography (CAG) results, 883 cases (506 control and 377 STEMI) from internal and external datasets were enrolled for testing DL models. Convolutional neural network-long short-term memory (CNN-LSTM) (AUC: 0.99) performed better than CNN, LSTM, and doctors in detecting STEMI. Deep learning models (AUC: 0.96) performed similarly to experienced cardiologists and emergency physicians in discriminating the left anterior descending (LAD) artery. Regarding distinguishing RCA from LCX, DL models were comparable to doctors (AUC: 0.81). In summary, we developed ECG-based DL diagnosis systems to detect STEMI and predict culprit vessel occlusion, thus enhancing the accuracy and effectiveness of STEMI diagnosis.

Deep learning assessment of left ventricular hypertrophy based on electrocardiogram.

Background: Current electrocardiogram (ECG) criteria of left ventricular hypertrophy (LVH) have low sensitivity. Deep learning (DL) techniques have been widely used to detect cardiac diseases due to its ability of automatic feature extraction of ECG. However, DL was rarely applied in LVH diagnosis. Our study aimed to construct a DL model for rapid and effective detection of LVH using 12-lead ECG. Methods: We built a DL model based on convolutional neural network-long short-term memory (CNN-LSTM) to detect LVH using 12-lead ECG. The echocardiogram and ECG of 1,863 patients obtained within 1 week after hospital admission were analyzed. Patients were evenly allocated into 3 sets at 3:1:1 ratio: the training set (n = 1,120), the validation set (n = 371) and the test set 1 (n = 372). In addition, we recruited 453 hospitalized patients into the internal test set 2. Different DL model of each subgroup was developed according to gender and relative wall thickness (RWT). Results: The LVH was predicted by the CNN-LSTM model with an area under the curve (AUC) of 0.62 (sensitivity 68%, specificity 57%) in the test set 1, which outperformed Cornell voltage criteria (AUC: 0.57, sensitivity 48%, specificity 72%) and Sokolow-Lyon voltage (AUC: 0.51, sensitivity 14%, specificity 96%). In the internal test set 2, the CNN-LSTM model had a stable performance in predicting LVH with an AUC of 0.59 (sensitivity 65%, specificity 57%). In the subgroup analysis, the CNN-LSTM model predicted LVH by 12-lead ECG with an AUC of 0.66 (sensitivity 72%, specificity 60%) for male patients, which performed better than that for female patients (AUC: 0.59, sensitivity 50%, specificity 71%). Conclusion: Our study established a CNN-LSTM model to diagnose LVH by 12-lead ECG with higher sensitivity than current ECG diagnostic criteria. This CNN-LSTM model may be a simple and effective screening tool of LVH.

Deep Learning Algorithms for the Prediction of Posttransplant Renal Function in Deceased-Donor Kidney Recipients: A Preliminary Study Based on Pretransplant Biopsy.

BACKGROUND: Posttransplant renal function is critically important for kidney transplant recipients. Accurate prediction of graft function would greatly help in deciding acceptance or discard of allocated kidneys. METHODS: : Whole-slide images (WSIs) of H&E-stained donor kidney biopsies at × 200 magnification between January 2015 and December 2019 were collected. The clinical characteristics of each donor and corresponding recipient were retrieved. Graft function was indexed with a stable estimated glomerular filtration rate (eGFR) and reduced graft function (RGF). We used convolutional neural network (CNN)-based models, such as EfficientNet-B5, Inception-V3, and VGG19 for the prediction of these two outcomes. RESULTS: In total, 219 recipients with H&E-stained slides of the donor kidneys were included for analysis [biopsies from standard criteria donor (SCD)/expanded criteria donor (ECD) was 191/28]. The results showed distinct improvements in the prediction performance of the deep learning algorithm plus the clinical characteristics model. The EfficientNet-B5 plus clinical data model showed the lowest mean absolute error (MAE) and root mean square error (RMSE). Compared with the clinical data model, the area under the receiver operating characteristic (ROC) curve (AUC) of the clinical data plus image model for eGFR classification increased from 0.69 to 0.83. In addition, the predictive performance for RGF increased from 0.66 to 0.80. Gradient-weighted class activation mappings (Grad-CAMs) showed that the models localized the areas of the tubules and interstitium near the glomeruli, which were discriminative features for RGF. CONCLUSION: Our results preliminarily show that deep learning for formalin-fixed paraffin-embedded H&E-stained WSIs improves graft function prediction accuracy for deceased-donor kidney transplant recipients.

Artificial intelligence for cellular phenotyping diagnosis of nasal polyps by whole-slide imaging.

BACKGROUND: artificial intelligence (AI) for cellular phenotyping diagnosis of nasal polyps by whole-slide imaging (WSI) is lacking. We aim to establish an AI chronic rhinosinusitis evaluation platform 2.0 (AICEP 2.0) to obtain the proportion of inflammatory cells for cellular phenotyping diagnosis of nasal polyps and to explore the clinical significance of different phenotypes of nasal polyps on the WSI. METHODS: a total of 453 patients were enrolled in our study. For the development of AICEP 2.0, 179 patients (WSIs) were obtained from the Third Affiliated Hospital of Sun Yat-Sen University (3HSYSU) from January 2008 to December 2018. A total of 24,625 patches were automatically extracted from the regions of interest under a 400× HPF by Openslide and the number of inflammatory cells in these patches was counted by two pathologists. For the application of AICEP 2.0 in a prospective cohort, 158 patients aged 14-70 years old with chronic rhinosinusitis with nasal polyps (CRSwNP) who had undergone endoscopic sinus surgery at 3HSYSU from June 2020 to December 2020 were included for preoperative demographic characteristics. For the application of AICEP 2.0 in a retrospective cohort, 116 patients with CRSwNP who had undergone endoscopic sinus surgery from May 2016 to June 2017 were enrolled for the recurrence rate. The proportion of inflammatory cells of these patients on WSI was calculated by our AICEP 2.0. FINDINGS: for AICEP 2.0, the mean absolute errors of the ratios of eosinophils, lymphocytes, neutrophils, and plasma cells were 1.64%, 2.13%, 1.06%, and 1.22%, respectively. The four phenotypes of nasal polyps were significantly different in clinical characteristics (including asthma, itching, sneezing, total IgE, peripheral eosinophils%, tissue eosinophils%, tissue neutrophils%, tissue lymphocytes%, tissue plasma cells%, and recurrence rate; P <0.05), but there were no significant differences in age distribution, onset time, total VAS score, Lund-Kennedy score, or Lund-Mackay score. The percentage of peripheral eosinophils was positively correlated with the percentage of tissue eosinophils (r = 0.560, P <0.001) and negatively correlated with tissue lymphocytes% (r = -0.489, P <0.001), tissue neutrophils% (r = -0.225, P = 0.005), and tissue plasma cells% (r = -0.266, P = 0.001) in WSIs.

A promising approach for screening pulmonary hypertension based on frontal chest radiographs using deep learning: A retrospective study.

BACKGROUND: To date, the missed diagnosis rate of pulmonary hypertension (PH) was high, and there has been limited development of a rapid, simple, and effective way to screen the disease. The purpose of this study is to develop a deep learning approach to achieve rapid detection of possible abnormalities in chest radiographs suggesting PH for screening patients suspected of PH. METHODS: We retrospectively collected frontal chest radiographs and the pulmonary artery systolic pressure (PASP) value measured by Doppler transthoracic echocardiography from 762 patients (357 healthy controls and 405 with PH) from three institutes in China from January 2013 to May 2019. The wohle sample comprised 762 images (641 for training, 80 for internal test, and 41 for external test). We firstly performed a 8-fold cross-validation on the 641 images selected for training (561 for pre-training, 80 for validation), then decided to tune learning rate to 0.0008 according to the best score on validation data. Finally, we used all the pre-training and validation data (561+80 = 641) to train our models (Resnet50, Xception, and Inception V3), evaluated them on internal and external test dataset to classify the images as having manifestations of PH or healthy according to the area under the receiver operating characteristic curve (AUC/ROC). After that, the three deep learning models were further used for prediction of PASP using regression algorithm. Moreover, we invited an experienced chest radiologist to classify the images in the test dataset as having PH or not, and compared the prediction accuracy performed by deep learing models with that of manual classification. RESULTS: The AUC performed by the best model (Inception V3) achieved 0.970 in the internal test, and slightly declined in the external test (0.967) when using deep learning algorithms to classify PH from normal based on chest X-rays. The mean absolute error (MAE) of the best model for prediction of PASP value was smaller in the internal test (7.45) compared to 9.95 in the external test. Manual classification of PH based on chest X-rays showed much lower AUCs compared to that performed by deep learning models both in the internal and external test. CONCLUSIONS: The present study used deep learning algorithms to classify abnormalities suggesting PH in chest radiographs with high accuracy and good generalizability. Once tested prospectively in clinical settings, the technology could provide a non-invasive and easy-to-use method to screen patients suspected of having PH.