Clot ratio, new clot burden score with deep learning, correlates with the risk stratification of patients with acute pulmonary embolism.

Background: Risk stratification for patients with acute pulmonary embolism (APE) is significantly important for treatment and prognosis evaluation. We aimed to develop a novel clot burden score on computed tomography pulmonary angiography (CTPA) based on deep learning (DL) algorithm for risk stratification of APE. Methods: The study retrospectively enrolled patients newly diagnosed with APE in China-Japan Friendship Hospital consecutively. We collected baseline data and CTPA parameters, and calculated four different clot burden scores, including Qanadli score, Mastora score, clot volume and clot ratio. The former two were calculated by two radiologists separately, while clot volume and clot ratio were based on the DL algorithm. The area under the curve (AUC) of four clot burden scores were analyzed. Results: Seventy patients were enrolled, including 17 in high-/intermediate-high risk and 53 in low-/intermediate-low risk. Clot burden was related to the risk stratification of APE. Among four clot burden scores, clot ratio had the highest AUC (0.719, 95% CI: 0.569-0.868) to predict patients with higher risk. In the patients with hemodynamically stable APE, only clot ratio presented statistical difference (P=0.046). Conclusions: Clot ratio is a new imaging marker of clot burden which correlates with the risk stratification of patients with APE. Higher clot ratio may indicate higher risk and acute right ventricular dysfunction in patients with hemodynamically stable status.

Development and validation of a lung graph-based machine learning model to predict acute pulmonary thromboembolism on chest noncontrast computed tomography.

Background: Computed tomography pulmonary angiography (CTPA) is a first-line noninvasive method to diagnose acute pulmonary thromboembolism (APE); however, whether chest noncontrast CT (NC-CT) could aid in the diagnosis of APE remains unknown. The aim of this study was to build and evaluate a holistic lung graph-based machine learning (HLG-ML) using NC-CT for the diagnosis of APE and to compare its performance with that of radiologists and the YEARS algorithm. Methods: This study enrolled 178 cases (77 males; age 63.9±16.7 years) who underwent NC-CT and CTPA in the same day from January 2019 to December 2020. Of these patients, 133 (75% of cases; 58 males; age 65.4±15.6 years) were placed into a training group and 45 (25% of cases; 19 males; age 59.6±19.2 years) into a testing group. The other 43 cases (18 males; age 62.8±20.0 years) were used to externally validate the model between January 2021 and March 2022. A HLG was developed with a pulmonary radiomics descriptor derived from NC-CT images. The approach extracted local radiomics features and encoded these local features into a radiomics descriptor as a characterization of global radiomics feature distribution. Subsequently, 8 ML models were trained and compared based on the radiomics descriptor. In the validation group, area under the curves (AUCs) of the HLG-ML model in the diagnosis of APE were compared with those of the 3 radiologists and the YEARS algorithm. Results: Among the 8 ML models, gradient boosting decision tree demonstrated the best classification performance (AUC =0.772) on the training set. In the testing set, the AUC of gradient boosting decision trees was 0.857 [95% confidence intervals (CIs): 0.699-0.951]. In the validation set, the performance of gradient boosting decision tree (AUC =0.810; 95% CI: 0.669-0.952; Youden index =0.621) outperformed 3 radiologists (AUC =0.508, 95% CI: 0.335-0.681, Youden index =0.016; AUC =0.504, 95% CI: 0.354-0.654, Youden index =0.008; AUC =0.527, 95% CI: 0.363-0.691, Youden index =0.050) and the YEARS algorithm (AUC =0.618; 95% CI: 0.469-0.767; Youden index =0.237). Conclusions: Compared to all 3 radiologists and the YEARS algorithm, the proposed HLG-based gradient boosting decision tree model achieved a superior performance in the diagnosis of APE on the NC-CT and may thus serve as a valuable tool for physicians in the diagnosis of APE.

Coronary computed tomography angiography analysis using artificial intelligence for stenosis quantification and stent segmentation: a multicenter study.

Background: Accurate interpretation of coronary computed tomography angiography (CCTA) is a labor-intensive and expertise-driven endeavor, as inexperienced readers may inadvertently overestimate stenosis severity. Recent artificial intelligence (AI) advances in medical imaging present compelling prospects for auxiliary diagnostic tools in CCTA. This study aimed to externally validate an AI-assisted analysis system capable of rapidly evaluating stenosis severity, exploring its potential integration into routine clinical workflows. Methods: This multicenter study consisted of an internal and external cohort of patients who underwent CCTA scans between April 2017 and February 2023. CCTA scans were evaluated using Coronary Artery Disease Reporting and Data System (CAD-RADS) scores to determine stenosis severity, while ground-truth stents were manually annotated by expert readers. The InferRead CT Heart (version 1.6; Infervision Medical Technology Co., Ltd., Beijing, China), which incorporates AI-assisted coronary artery stenosis quantification and automatic stent segmentation, was employed for CCTA scan analysis. AI-based stenosis assessment performance was determined using sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), while the AI-based stent segmentation overlap was assessed using the Dice similarity coefficient (DSC). Results: For ≥50% stenosis diagnoses, the AI system attained per-patient sensitivity, specificity, PPV, and NPV surpassing 90.0% for the internal dataset; for the external dataset, the per-patient values were 88.0% [95% confidence interval (CI): 81.0-94.4%], 94.5% (95% CI: 90.7-97.6%), 90.0% (95% CI: 83.3-95.6%), and 93.4% (95% CI: 89.2-96.8%), respectively. For ≥70% stenosis diagnoses, the per-patient values on the internal dataset were 94.2% (95% CI: 89.2-98.1%), 95.8% (95% CI: 94.1-97.4%), 80.8% (95% CI: 73.5-87.7%), and 98.9% (95% CI: 97.9-99.6%), respectively; for the external dataset, the per-patient values were 91.9% (95% CI: 82.6-100.0%), 97.3% (95% CI: 94.9-99.1%), 85.0% (95% CI: 72.5-94.6%), and 98.6% (95% CI: 96.8-100.0%), respectively. Regarding CAD-RADS categorization, the Cohen kappa was 0.75 and 0.81 for the internal per-patient and per-vessel basis, respectively, and 0.72 and 0.76 for the external per-patient and per-vessel basis, respectively. The DSC for stent segmentation was 0.96±0.06. Conclusions: The AI-assisted analysis system for CCTA interpretation exhibited exceptional proficiency in stenosis quantification and stent segmentation, indicating that AI holds considerable potential in advancing CCTA postprocessing techniques.

Idiopathic pulmonary fibrosis disease progression: a dynamic quantitative chest computed tomography follow-up analysis.

Background: To clarify whether dynamic quantification of variables derived from chest high-resolution computed tomography (HRCT) can assess the progression of idiopathic pulmonary fibrosis (IPF). Methods: Patients with IPF who underwent serial computed tomography (CT) imaging were retrospectively enrolled. Several structural abnormalities seen on HRCT in IPF were segmented and quantified. Patients were divided into 2 groups according to their pulmonary function test (PFT) results: those with disease stabilization and those with disease progression, and differences between the groups were analyzed. Results: There were no statistically significant differences between the 2 patient groups for the following parameters: baseline PFTs, total lesion extent, lesion extent at different sites in the lungs, and pulmonary vessel-related parameters (with P values ranging from 0.057 to 0.894). Median changes in total lung volume, total lesion volume, and total lesion ratio were significantly higher in patients with worsening disease compared with those with stable disease (P<0.001). There was a significant increase in total lesion volume of 214.73 mL [interquartile range (IQR), 68.26 to 501.46 mL] compared with 3.67 mL (IQR, -71.70 to 85.33 mL) in the disease progression group compared with the disease stability group (P=0.001). The decline in pulmonary vessel volume and number of pulmonary vessel branches was more pronounced in the group with functional worsening compared with the group with functional stability. Moreover, changes in lesion volume ratio were negatively correlated with changes in diffusing capacity of the lungs for carbon monoxide (DLco) during follow-up (R=-0.57, P<0.001), and changes in pulmonary vessel-related parameters demonstrated positive correlation with DLco (with R ranging from 0.27 to 0.53, P<0.001) and forced vital capacity (FVC) (with R ranging from 0.44 to 0.61, P<0.001). Conclusions: Changes in CT-related parameters during follow-up may have better predictive performance compared with baseline imaging parameters and PFTs for disease progression in IPF.

Lung shrinking assessment on HRCT with elastic registration technique for monitoring idiopathic pulmonary fibrosis.

OBJECTIVES: Evaluation and follow-up of idiopathic pulmonary fibrosis (IPF) mainly rely on high-resolution computed tomography (HRCT) and pulmonary function tests (PFTs). The elastic registration technique can quantitatively assess lung shrinkage. We aimed to investigate the correlation between lung shrinkage and morphological and functional deterioration in IPF. METHODS: Patients with IPF who underwent at least two HRCT scans and PFTs were retrospectively included. Elastic registration was performed on the baseline and follow-up HRCTs to obtain deformation maps of the whole lung. Jacobian determinants were calculated from the deformation fields and after logarithm transformation, log_jac values were represented on color maps to describe morphological deterioration, and to assess the correlation between log_jac values and PFTs. RESULTS: A total of 69 patients with IPF (male 66) were included. Jacobian maps demonstrated constriction of the lung parenchyma marked at the lung base in patients who were deteriorated on visual and PFT assessment. The log_jac values were significantly reduced in the deteriorated patients compared to the stable patients. Mean log_jac values showed positive correlation with baseline percentage of predicted vital capacity (VC%) (r = 0.394, p < 0.05) and percentage of predicted forced vital capacity (FVC%) (r = 0.395, p < 0.05). Additionally, the mean log_jac values were positively correlated with pulmonary vascular volume (r = 0.438, p < 0.01) and the number of pulmonary vascular branches (r = 0.326, p < 0.01). CONCLUSIONS: Elastic registration between baseline and follow-up HRCT was helpful to quantitatively assess the morphological deterioration of lung shrinkage in IPF, and the quantitative indicator log_jac values were significantly correlated with PFTs. KEY POINTS: • The elastic registration on HRCT was helpful to quantitatively assess the deterioration of IPF. • Jacobian logarithm was significantly reduced in deteriorated patients and mean log_jac values were correlated with PFTs. • The mean log_jac values were related to the changes of pulmonary vascular volume and the number of vascular branches.

Characteristics of Diaphragmatic and Chest Wall Motion in People with Normal Pulmonary Function: A Study with Free-Breathing Dynamic MRI.

Objective: We aimed to quantitatively study the characteristic of diaphragm and chest wall motion using free-breathing dynamic magnetic resonance imaging (D-MRI) in Chinese people with normal lung function. Methods: 74 male subjects (mean age, 37 ± 11 years old) were prospectively enrolled, and they underwent high-resolution CT(HRCT), pulmonary functional tests (PFTs), and D-MRI in the same day. D-MRI was acquired with a gradient-echo sequence during the quiet and deep breathing. The motion of the diaphragm and chest wall were respectively assessed by measuring thoracic anteroposterior diameter (AP), left−right diameter (LR), cranial−caudal diameter (CC), and thoracic area ratios between end-inspiration and end-expiration. The effect of age, body mass index (BMI), and smoking on respiratory muscle function was also analyzed. Results: The mean ratio of right and left AP was greater than that of LR on three transversal planes during both quiet and deep breathing. The mean ratio at the anterior diaphragm (AND, Quiet: 1.04 ± 0.03; Deep: 1.15 ± 0.09) was weaker than that of the apex (vs. APD, Quiet: 1.08 ± 0.05, p < 0.001; Deep: 1.29 ± 0.12, p < 0.001) and posterior diaphragm (vs. POD, Quiet: 1.09 ± 0.04, p < 0.001; Deep: 1.30 ± 0.12, p < 0.001) both in quiet and deep breathing. Compared with non-smokers, the left AP and thoracic area ratios in smokers were significantly decreased (p < 0.05). However, the ratios of AP, LR, CC, and thoracic area on each plane were similar among groups in different age and BMI. Conclusions: During both quiet and deep breathing, the chest wall motion is prominent in the anteroposterior direction. The motions of diaphragm apex and posterior diaphragm were more prominent than that of the anterior diaphragm. Smoking may affect the respiratory muscle mobility. Dynamic MRI can quantitatively evaluate the motion of respiratory muscles.

MMMNA-Net for Overall Survival Time Prediction of Brain Tumor Patients.

Overall survival (OS) time is one of the most important evaluation indices for gliomas situations. Multi-modal Magnetic Resonance Imaging (MRI) scans play an important role in the study of glioma prognosis OS time. Several deep learning-based methods are proposed for the OS time prediction on multi-modal MRI problems. However, these methods usually fuse multi-modal information at the beginning or at the end of the deep learning networks and lack the fusion of features from different scales. In addition, the fusion at the end of networks always adapts global with global (eg. fully connected after concatenation of global average pooling output) or local with local (eg. bilinear pooling), which loses the information of local with global. In this paper, we propose a novel method for multi-modal OS time prediction of brain tumor patients, which contains an improved non-local features fusion module introduced on different scales. Our method obtains a relative 8.76% improvement over the current state-of-art method (0.6989 vs. 0.6426 on accuracy). An extra testing demonstrates that our method could adapt to the situations with missing modalities. The code is available at https://github.com/TangWen920812/mmmna-net.

Quantitative analysis of high-resolution computed tomography features of idiopathic pulmonary fibrosis: a structure-function correlation study.

Background: The quantitative analysis of high-resolution computed tomography (HRCT) is increasingly being used to quantify the severity and evaluate the prognosis of disease. Our aim was to quantify the HRCT features of idiopathic pulmonary fibrosis (IPF) and identify their association with pulmonary function tests. Methods: This was a retrospective, single-center, clinical research study. Patients with IPF were retrospectively included. Pulmonary segmentation was performed using the deep learning-based method. Radiologists manually segmented 4 findings of IPF, including honeycombing (HC), reticular pattern (RE), traction bronchiectasis (TRBR), and ground glass opacity (GGO). Pulmonary vessels were segmented with the automatic integration segmentation method. All segmentation results were quantified by the corresponding segmentation software. Correlations between the volume of the 4 findings on HRCT, volume of the lesions at different sites, pulmonary vascular-related parameters, and pulmonary function tests were analyzed. Results: A total of 101 IPF patients (93 males) with a median age of 63 years [interquartile range (IQR), 58 to 68 years] were included in this study. Total lesion extent demonstrated a stronger negative correlation with diffusion capacity for carbon monoxide (DLco) compared to HC, RE, and TRBR [total lesion ratio, correlation coefficient (r) =-0.67, P<0.001; HC, r=-0.45, P<0.001; RE, r=-0.41, P<0.001; TRBR, r=-0.25, P<0.05, respectively]. Correlations with lung function were similar among various lesion sites with r from -0.38 to -0.61 (P<0.001). Pulmonary artery volume (PAV) displayed a slightly increased positive association with the DLco compared to total pulmonary vascular volume (PVV); for PAV, r=0.41 and P<0.001 and for total PVV, r=0.36 and P<0.001. Additionally, total lesion extent, HC, and RE indicated a negative relationship with vascular-related parameters, and the strength of the correlations was independent of lesion site. Conclusions: Quantitative analysis of HRCT features of IPF indicated a decline in function and an aggravation of vascular destruction with increasing lesion extent. Furthermore, a positive correlation between vascular-related parameters and pulmonary function was confirmed. This co-linearity indicated the potential of vascular-related parameters as new objective markers for evaluating the severity of IPF.

Non-invasive precise staging of liver fibrosis using deep residual network model based on plain CT images.

PURPOSE: The aim of this study was to explore the application of five-class deep residual network models based on plain CT images and clinical features for the precise staging of liver fibrosis. METHODS: This retrospective clinical study included 347 patients who underwent liver CT, with pathological staging of liver fibrosis as the gold standard. We established three ResNet models to stage liver fibrosis. The output diagnosis labels of models were 0, 1, 2, 3 and 4, which correspond to F0, F1, F2, F3, and F4 stages. Confusion matrices were used to evaluate the performances of models to precisely stage liver fibrosis. The performance for diagnosing cirrhosis (F4), advanced fibrosis (≥ F3) and significant fibrosis (≥ F2) of models was evaluated with receiver operating characteristic (ROC) analyses. RESULTS: The kappa coefficients of the five-class ResNet model (based on plain CT images), the five-class ResNet clinical model (based on clinical features), and the five-class mixed ResNet model (based on plain CT images and clinical features) for precise staging liver fibrosis were 0.566, 0.306, and 0.63, respectively. The recall rates and precision rates for F0, F1, F2, and F3 of three models were lower than 60%. The ROC AUC values of the five-class ResNet model, the five-class ResNet clinical model, and the five-class mixed ResNet model for diagnosing cirrhosis, advanced fibrosis, and significant fibrosis were 0.95, 0.88, and 0.82, 0.80, 0.72, and 0.70, 0.95, 0.90, and 0.83, respectively. CONCLUSIONS: The five-class ResNet models are of high value in the diagnosis of liver cirrhosis, advanced liver fibrosis, and significant liver fibrosis. However, for the precise staging of liver fibrosis, the models cannot accurately distinguish other liver fibrosis stages except F4. Plain CT images combined with clinical features have the potential to improve the performance of the ResNet models in diagnosing liver fibrosis.

Clot burden of acute pulmonary thromboembolism: comparison of two deep learning algorithms, Qanadli score, and Mastora score.

BACKGROUND: The deep learning convolution neural network (DL-CNN) benefits evaluating clot burden of acute pulmonary thromboembolism (APE). Our objective was to compare the performance of the deep learning convolution neural network trained by the fine-tuning [DL-CNN (ft)] and the deep learning convolution neural network trained from the scratch [DL-CNN (fs)] in the quantitative assessment of APE. METHODS: We included the data of 680 cases for training DL-CNN by DL-CNN (ft) and DL-CNN (fs), then retrospectively included 410 patients (137 patients with APE, 203 males, mean age 60.3±11.4 years) for testing the models. The distribution and volume of clots were respectively assessed by DL-CNN(ft) and DL-CNN(fs), and sensitivity, specificity, and area under the curve (AUC) were used to evaluate their performances in detecting clots on a per-patient and clot level. Radiologists evaluated the distribution of clots, Qanadli score, and Mastora score and right ventricular metrics, and the correlation of clot volumes with right ventricular metrics were analyzed with Spearman correlation analysis. RESULTS: On a per-patient level, the two DL-CNN models had high sensitivities and moderate specificities [DL-CNN (ft): 100% and 77.29%; DL-CNN (fs): 100% and 75.82%], and their AUCs were comparable (Z=0.30, P=0.38). On a clot level, DL-CNN (ft) and DL-CNN (fs) sensitivities and specificities in detecting central clots were 99.06% and 72.61%, and 100% and 70.63%, respectively. DL-CNN (ft) sensitivities and specificities in detecting peripheral clots were mostly higher than those of DL-CNN (fs), and their AUCs were comparable. Clot volumes measured with the two models were similar (U=85094.500, P=0.741), and significantly correlated with Qanadli scores [DL-CNN(ft) r=0.825, P<0.001, DL-CNN(fs) r=0.827, P<0.001] and Mastora scores [DL-CNN(ft) r=0.859, P<0.001, DL-CNN(fs) r=0.864, P<0.001]. Clot volumes were also correlated with right ventricular metrics. Clot burdens were increased in the low-risk, moderate-risk, and high-risk patients. Binary logistic regression revealed that only the ratio of right ventricular area/left ventricular area (RVa/LVa) was an independent predictor of in-hospital death (odds ratio 6.73; 95% CI, 2.7-18.12, P<0.001). CONCLUSIONS: Both DL-CNN (ft) and DL-CNN (fs) have high sensitivities and moderate specificities in detecting clots associated with APE, and their performances are comparable. While clot burdens quantitatively calculated by the two DL-CNN models are correlated with right ventricular function and risk stratification, RVa/LVa is an independent prognostic factor of in-hospital death in patients with APE.

Machine learning analysis for the noninvasive prediction of lymphovascular invasion in gastric cancer using PET/CT and enhanced CT-based radiomics and clinical variables.

PURPOSE: Lymphovascular invasion (LVI) is associated with metastasis and poor survival in patients with gastric cancer, yet the noninvasive diagnosis of LVI is difficult. This study aims to develop predictive models using different machine learning (ML) classifiers based on both enhanced CT and PET/CT images and clinical variables for preoperatively predicting lymphovascular invasion (LVI) status of gastric cancer. METHODS: A total of 101 patients with gastric cancer who underwent surgery were retrospectively recruited, and the LVI status was confirmed by pathological analysis. Patients were randomly divided into a training dataset (n = 76) and a validation dataset (n = 25). By 3D manual segmentation, radiomics features were extracted from the PET and venous phase CT images. Image models, clinical models, and combined models were constructed by selected enhanced CT-based and PET-based radiomics features, clinical factors, and a combination of both, respectively. Three ML classifiers including adaptive boosting (AdaBoost), linear discriminant analysis (LDA), and logistic regression (LR) were used for model development. The performance of these predictive models was evaluated with respect to discrimination, calibration, and clinical usefulness. RESULTS: Ten radiomics features and eight clinical factors were selected for the development of predictive models. In the validation dataset, the area under curve (AUC) values of clinical models using AdaBoost, LDA, and LR classifiers were 0.742, 0.706, and 0.690, respectively. The image models using AdaBoost, LDA, and LR classifiers achieved an AUC of 0.849, 0.778, and 0.810, respectively. The combined models showed improved performance than the image models and the clinical models, with the AUC values of AdaBoost, LDA, and LR classifier yielding 0.944, 0.929, and 0.921, respectively. The combined models also showed good calibration and clinical usefulness for LVI prediction. CONCLUSION: ML-based models integrating PET/CT and enhanced CT radiomics features and clinical factors have good discrimination capability, which could serve as a noninvasive, preoperative tool for the prediction of LVI and assist surgical treatment decisions in patients with gastric cancer.

Using contrast-enhanced CT and non-contrast-enhanced CT to predict EGFR mutation status in NSCLC patients-a radiomics nomogram analysis.

OBJECTIVES: To develop and validate a general radiomics nomogram capable of identifying EGFR mutation status in non-small cell lung cancer (NSCLC) patients, regardless of patient with either contrast-enhanced CT (CE-CT) or non-contrast-enhanced CT (NE-CT). METHODS: A total of 412 NSCLC patients were retrospectively enrolled in this study. Patients' radiomics features not significantly different between NE-CT and CE-CT were defined as general features, and were further used to construct the general radiomics signature. Fivefold cross-validation was used to select the best machine learning algorithm. Finally, a general radiomics nomogram was developed using general radiomics signature, and clinical and radiological characteristics. Two groups of data collected at different time periods were used as two test sets to access the discrimination and clinical usefulness. Area under the receiver operating characteristic curve (ROC-AUC) was applied to performance evaluation. RESULT: The general radiomics signature yielded the highest AUC of 0.756 and 0.739 in the two test sets, respectively. When applying to same type of CT, the performance of general radiomics signature was always similar to or higher than that of models built using only NE-CT or CE-CT features. The general radiomics nomogram combining general radiomics signature, smoking history, emphysema, and ILD achieved higher performance whether applying to NE-CT or CE-CT (test set 1, AUC = 0.833 and 0.842; test set 2, AUC = 0.839 and 0.850). CONCLUSIONS: Our work demonstrated that using general features to construct radiomics signature and nomogram could help identify EGFR mutation status of NSCLC patients and expand its scope of clinical application. KEY POINTS: • General features were proposed to construct general radiomics signature using different types of CT of different patients at the same time to identify EGFR mutation status of NSCLC patients. • The general radiomics nomogram based on general radiomics signature, and clinical and radiological characteristics could identify EGFR mutation status of patients with NSCLC and outperformed the general radiomics signature. • The general radiomics nomogram had a wider scope of clinical application; no matter which of NE-CT and CE-CT the patient has, its EGFR mutation status could be predicted.

Assessing the predictive accuracy of lung cancer, metastases, and benign lesions using an artificial intelligence-driven computer aided diagnosis system.

BACKGROUND: Artificial intelligence (AI) products have been widely used for the clinical detection of primary lung tumors. However, their performance and accuracy in risk prediction for metastases or benign lesions remain underexplored. This study evaluated the accuracy of an AI-driven commercial computer-aided detection (CAD) product (InferRead CT Lung Research, ICLR) in malignancy risk prediction using a real-world database. METHODS: This retrospective study assessed 486 consecutive resected lung lesions, including 320 adenocarcinomas, 40 other malignancies, 55 metastases, and 71 benign lesions, from September 2015 to November 2018. The malignancy risk probability of each lesion was obtained using the ICLR software based on a 3D convolutional neural network (CNN) with DenseNet architecture as a backbone (without clinical data). Two resident doctors independently graded each lesion using patient clinical history. One doctor (R1) has 3 years of chest radiology experience, and the other doctor (R2) has 3 years of general radiology experience. Cochran's Q test was used to assess the performances of the AI compared to the radiologists. RESULTS: The accuracy of malignancy-risk prediction using the ICLR for adenocarcinomas, other malignancies, metastases, and benign lesions was 93.4% (299/320), 95.0% (38/40), 50.9% (28/55), and 40.8% (29/71), respectively. The accuracy was significantly higher in adenocarcinomas and other malignancies compared to metastases and benign lesions (all P<0.05). The overall accuracy of risk prediction for R1 was 93.6% (455/486) and 87.4% for R2 (425/486), both of which were higher than the 81.1% accuracy obtained with the ICLR (394/486) (R1 vs. ICLR: P<0.001; R2 vs. ICLR: P=0.001), especially in assessing the risk of metastases (P<0.05). R1 performed better than R2 at risk prediction (P=0.001). CONCLUSIONS: The accuracy of the ICLR for risk prediction is very high for primary lung cancers but poor for metastases and benign lesions.

3D DenseNet Deep Learning Based Preoperative Computed Tomography for Detecting Myasthenia Gravis in Patients With Thymoma.

BACKGROUND: Myasthenia gravis (MG) is the most common paraneoplastic syndromes of thymoma and closely related to thymus abnormalities. Timely detecting of the risk of MG would benefit clinical management and treatment decision for patients with thymoma. Herein, we developed a 3D DenseNet deep learning (DL) model based on preoperative computed tomography (CT) as a non-invasive method to detect MG in thymoma patients. METHODS: A large cohort of 230 thymoma patients in a hospital affiliated with a medical school were enrolled. 182 thymoma patients (81 with MG, 101 without MG) were used for training and model building. 48 cases from another hospital were used for external validation. A 3D-DenseNet-DL model and five radiomic models were performed to detect MG in thymoma patients. A comprehensive analysis by integrating machine learning and semantic CT image features, named 3D-DenseNet-DL-based multi-model, was also performed to establish a more effective prediction model. FINDINGS: By elaborately comparing the prediction efficacy, the 3D-DenseNet-DL effectively identified MG patients and was superior to other five radiomic models, with a mean area under ROC curve (AUC), accuracy, sensitivity, and specificity of 0.734, 0.724, 0.787, and 0.672, respectively. The effectiveness of the 3D-DenseNet-DL-based multi-model was further improved as evidenced by the following metrics: AUC 0.766, accuracy 0.790, sensitivity 0.739, and specificity 0.801. External verification results confirmed the feasibility of this DL-based multi-model with metrics: AUC 0.730, accuracy 0.732, sensitivity 0.700, and specificity 0.690, respectively. INTERPRETATION: Our 3D-DenseNet-DL model can effectively detect MG in patients with thymoma based on preoperative CT imaging. This model may serve as a supplement to the conventional diagnostic criteria for identifying thymoma associated MG.

Can peritumoral regions increase the efficiency of machine-learning prediction of pathological invasiveness in lung adenocarcinoma manifesting as ground-glass nodules?

BACKGROUND: The peri-tumor microenvironment plays an important role in the occurrence, growth and metastasis of cancer. The aim of this study is to explore the value and application of a CT image-based deep learning model of tumors and peri-tumors in predicting the invasiveness of ground-glass nodules (GGNs). METHODS: Preoperative thin-section chest CT images were reviewed retrospectively in 622 patients with a total of 687 pulmonary GGNs. GGNs are classified according to clinical management strategies as invasive lesions (IAC) and non-invasive lesions (AAH, AIS and MIA). The two volumes of interest (VOIs) identified on CT were the gross tumor volume (GTV) and the gross volume of tumor incorporating peritumoral region (GPTV). Three dimensional (3D) DenseNet was used to model and predict GGN invasiveness, and five-fold cross validation was performed. We used GTV and GPTV as inputs for the comparison model. Prediction performance was evaluated by sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). RESULTS: The GTV-based model was able to successfully predict GGN invasiveness, with an AUC of 0.921 (95% CI, 0.896-0.937). Using GPTV, the AUC of the model increased to 0.955 (95% CI, 0.939-0.971). CONCLUSIONS: The deep learning method performed well in predicting GGN invasiveness. The predictive ability of the GPTV-based model was more effective than that of the GTV-based model.

Opportunistic osteoporosis screening in multi-detector CT images using deep convolutional neural networks.

OBJECTIVE: To explore the application of deep learning in patients with primary osteoporosis, and to develop a fully automatic method based on deep convolutional neural network (DCNN) for vertebral body segmentation and bone mineral density (BMD) calculation in CT images. MATERIALS AND METHODS: A total of 1449 patients were used for experiments and analysis in this retrospective study, who underwent spinal or abdominal CT scans for other indications between March 2018 and May 2020. All data was gathered from three different CT vendors. Among them, 586 cases were used for training, and other 863 cases were used for testing. A fully convolutional neural network, called U-Net, was employed for automated vertebral body segmentation. The manually sketched region of vertebral body was used as the ground truth for comparison. A convolutional neural network, called DenseNet-121, was applied for BMD calculation. The values post-processed by quantitative computed tomography (QCT) were identified as the standards for analysis. RESULTS: Based on the diversity of CT vendors, all testing cases were split into three testing cohorts: Test set 1 (n = 463), test set 2 (n = 200), and test set 3 (n = 200). Automated segmentation correlated well with manual segmentation regarding four lumbar vertebral bodies (L1-L4): the minimum average dice coefficients for three testing sets were 0.823, 0.786, and 0.782, respectively. For testing sets from different vendors, the average BMDs calculated by automated regression showed high correlation (r > 0.98) and agreement with those derived from QCT. CONCLUSIONS: A deep learning-based method could achieve fully automatic identification of osteoporosis, osteopenia, and normal bone mineral density in CT images. KEY POINTS: • Deep learning can perform accurate fully automated segmentation of lumbar vertebral body in CT images. • The average BMDs obtained by deep learning highly correlates with ones derived from QCT. • The deep learning-based method could be helpful for clinicians in opportunistic osteoporosis screening in spinal or abdominal CT scans.

[Prediction of Pathological Subtypes of Lung Adenocarcinoma with Pure Ground Glass Nodules by Deep Learning Model].

Objective To make a preliminary pathological classification of lung adenocarcinoma with pure ground glass nodules(pGGN)on CT by using a deep learning model. Methods CT images and pathological data of 219 patients(240 lesions in total)with pGGN on CT and pathologically confirmed adenocarcinoma were collected.According to pathological subtypes,the lesions were divided into non-invasive lung adenocarcinoma group(which included atypical adenomatous hyperplasia and adenocarcinoma in situ and micro-invasive adenocarcinoma)and invasive lung adenocarcinoma group.First,the lesions were outlined and labeled by two young radiologists,and then the labeled data were randomly divided into two datasets:the training set(80%)and the test set(20%).The prediction Results of deep learning were compared with those of two experienced radiologists by using the test dataset. Results The deep learning model achieved high performance in predicting the pathological types(non-invasive and invasive)of pGGN lung adenocarcinoma.The accuracy rate in pGGN diagnosis was 0.8330(95% CI=0.7016-0.9157)for of deep learning model,0.5000(95% CI=0.3639-0.6361)for expert 1,0.5625(95% CI=0.4227-0.6931)for expert 2,and 0.5417(95% CI=0.4029-0.6743)for both two experts.Thus,the accuracy of the deep learning model was significantly higher than those of the experienced radiologists(P=0.002).The intra-observer agreements were good(Kappa values:0.939 and 0.799,respectively).The inter-observer agreement was general(Kappa value:0.667)(P=0.000). Conclusion The deep learning model showed better performance in predicting the pathological types of pGGN lung adenocarcinoma compared with experienced radiologists.

Deep residual nets model for staging liver fibrosis on plain CT images.

PURPOSE: The early diagnosis of liver fibrosis is crucial for the prevention of liver cirrhosis and liver cancer. As gold standard for staging liver fibrosis, liver biopsy is an invasive procedure that carries the risk of serious complications. The aim of this study was to evaluate the performance of the residual neural network (ResNet), a non-invasive methods, for staging liver fibrosis using plain CT images. METHODS: This retrospective study involved 347 patients subjected to liver CT scanning and liver biopsy. For each patient, we selected three axial images adjacent to the puncture location in the eighth or ninth inter-space on the right side. After processing and enhancement (rotation, translation, and amplification), these images were used as input data for the ResNet model. The model used a fivefold cross-validation method. In each fold, the images of approximately 80% of the total sample size (278 patients) were used for training the ResNet model, the other 20% (69 patients) were used for testing the trained network, with the liver biopsy pathology results as gold standard. The proportion of patients in each fibrosis stage was equal for training and test groups. The final result was the mean of the fivefold cross-validation in the test group. The performance of the ResNet model was evaluated for the test group by receiver operating characteristic (ROC) analysis. RESULTS: For the ResNet model, the area under the ROC curve (AUC) for assessing cirrhosis (F4), advanced fibrosis (F3 or higher), significant fibrosis (F2 or higher), and mild fibrosis (F1 or higher) was 0.97, 0.94, 0.90, and 0.91, respectively. CONCLUSIONS: The ResNet model analysis of plain CT images exhibited high diagnostic efficiency for liver fibrosis staging. As a convenient, fast, and economical non-invasive diagnostic method, the ResNet model can be used to assist radiologists and clinicians in liver fibrosis evaluations.

Identification of benign and malignant pulmonary nodules on chest CT using improved 3D U-Net deep learning framework.

PURPOSE: To accurately distinguish benign from malignant pulmonary nodules with CT based on partial structures of 3D U-Net integrated with Capsule Networks (CapNets) and provide a reference for the early diagnosis of lung cancer. METHOD: The dataset consisted of 1177 samples (benign/malignant: 414/763) from 997 patients provided by collaborating hospital. All nodules were biopsy or surgery proven, and pathologic results were regarded as the "golden standard". This study utilized partial U-Net to capture the low-level (edge, corner, etc.) information and CapNets to preserve high-level (semantic information) information of nodules. For CapNets, each capsule had a 4 × 4 matrix representing the pose and an activation probability representing the presence of an object. Furthermore, we chose accuracy (ACC), area under the curve (AUC), sensitivity (SE) and specificity (SP) to evaluate the generalization of the proposed architecture and compared its identification performance with 3D U-Net and experienced radiologists. RESULTS: The AUC of our architecture (0.84) was superior to that (0.81) of the original 3D U-Net (p = 0.04, DeLong's test). Moreover, ACC (84.5 %) and SE (92.9 %) of our model were clearly higher than radiologists' ACC (81.0 %) and SE (84.3 %) at the optimal operating point. However, SP (70 %) of our model was slightly lower than radiologists' SP (75 %), which might be the result of class imbalance with limited benign samples involved for algorithm training. CONCLUSIONS: Our architecture showed a high performance for identifying benign and malignant pulmonary nodules, indicating the improved model has a promising application in clinic.

Development of a deep learning model for classifying thymoma as Masaoka-Koga stage I or II via preoperative CT images.

BACKGROUND: Accurate thymoma staging via computed tomography (CT) images is difficult even for experienced thoracic doctors. Here we developed a preoperative staging tool differentiating Masaoka-Koga (MK) stage I patients from stage II patients using CT images. METHODS: CT images of 174 thymoma patients were retrospectively selected. Two chest radiologists independently assessed the images. Variables with statistical differences in univariate analysis were adjusted for age, sex, and smoking history in multivariate logical regression to determine independent predictors of the thymoma stage. We established a deep learning (DL) 3D-DenseNet model to distinguish the MK stage I and stage II thymomas. Furthermore, we compared two different methods to label the regions of interest (ROI) in CT images. RESULTS: In routine CT images, there were statistical differences (P<0.05) in contour, necrosis, cystic components, and the degree of enhancement between stage I and II disease. Multivariate logical regression showed that only the degree of enhancement was an independent predictor of the thymoma stage. The area under the receiver operating characteristic curve (AUC) of routine CT images for classifying thymoma as MK stage I or II was low (AUC =0.639). The AUC of the 3D-DenseNet model showed better performance with a higher AUC (0.773). ROIs outlined by segmentation labels performed better (AUC =0.773) than those outlined by bounding box labels (AUC =0.722). CONCLUSIONS: Our DL 3D-DenseNet may aid thymoma stage classification, which may ultimately guide surgical treatment and improve outcomes. Compared with conventional methods, this approach provides improved staging accuracy. Moreover, ROIs labeled by segmentation is more recommendable when the sample size is limited.