Predicting Invasiveness of Lung Adenocarcinoma at Chest CT with Deep Learning Ternary Classification Models.

Background Preoperative discrimination of preinvasive, minimally invasive, and invasive adenocarcinoma at CT informs clinical management decisions but may be challenging for classifying pure ground-glass nodules (pGGNs). Deep learning (DL) may improve ternary classification. Purpose To determine whether a strategy that includes an adjudication approach can enhance the performance of DL ternary classification models in predicting the invasiveness of adenocarcinoma at chest CT and maintain performance in classifying pGGNs. Materials and Methods In this retrospective study, six ternary models for classifying preinvasive, minimally invasive, and invasive adenocarcinoma were developed using a multicenter data set of lung nodules. The DL-based models were progressively modified through framework optimization, joint learning, and an adjudication strategy (simulating a multireader approach to resolving discordant nodule classifications), integrating two binary classification models with a ternary classification model to resolve discordant classifications sequentially. The six ternary models were then tested on an external data set of pGGNs imaged between December 2019 and January 2021. Diagnostic performance including accuracy, specificity, and sensitivity was assessed. The χ2 test was used to compare model performance in different subgroups stratified by clinical confounders. Results A total of 4929 nodules from 4483 patients (mean age, 50.1 years ± 9.5 [SD]; 2806 female) were divided into training (n = 3384), validation (n = 579), and internal (n = 966) test sets. A total of 361 pGGNs from 281 patients (mean age, 55.2 years ± 11.1 [SD]; 186 female) formed the external test set. The proposed strategy improved DL model performance in external testing (P < .001). For classifying minimally invasive adenocarcinoma, the accuracy was 85% and 79%, sensitivity was 75% and 63%, and specificity was 89% and 85% for the model with adjudication (model 6) and the model without (model 3), respectively. Model 6 showed a relatively narrow range (maximum minus minimum) across diagnostic indexes (accuracy, 1.7%; sensitivity, 7.3%; specificity, 0.9%) compared with the other models (accuracy, 0.6%-10.8%; sensitivity, 14%-39.1%; specificity, 5.5%-17.9%). Conclusion Combining framework optimization, joint learning, and an adjudication approach improved DL classification of adenocarcinoma invasiveness at chest CT. Published under a CC BY 4.0 license. Supplemental material is available for this article. See also the editorial by Sohn and Fields in this issue.

[CT-Based Weighted Radiomic Score Predicts Tumor Response to Immunotherapy in Non-Small Cell Lung Cancer].

Objective To develop a CT-based weighted radiomic model that predicts tumor response to programmed death-1(PD-1)/PD-ligand 1(PD-L1)immunotherapy in patients with non-small cell lung cancer.Methods The patients with non-small cell lung cancer treated by PD-1/PD-L1 immune checkpoint inhibitors in the Peking Union Medical College Hospital from June 2015 to February 2022 were retrospectively studied and classified as responders(partial or complete response)and non-responders(stable or progressive disease).Original radiomic features were extracted from multiple intrapulmonary lesions in the contrast-enhanced CT scans of the arterial phase,and then weighted and summed by an attention-based multiple instances learning algorithm.Logistic regression was employed to build a weighted radiomic scoring model and the radiomic score was then calculated.The area under the receiver operating characteristic curve(AUC)was used to compare the weighted radiomic scoring model,PD-L1 model,clinical model,weighted radiomic scoring + PD-L1 model,and comprehensive prediction model.Results A total of 237 patients were included in the study and randomized into a training set(n=165)and a test set(n=72),with the mean ages of(64±9)and(62±8)years,respectively.The AUC of the weighted radiomic scoring model reached 0.85 and 0.80 in the training set and test set,respectively,which was higher than that of the PD-L1-1 model(Z=37.30,P<0.001 and Z=5.69,P=0.017),PD-L1-50 model(Z=38.36,P<0.001 and Z=17.99,P<0.001),and clinical model(Z=11.40,P<0.001 and Z=5.76,P=0.016).The AUC of the weighted scoring model was not different from that of the weighted radiomic scoring + PD-L1 model and the comprehensive prediction model(both P>0.05).Conclusion The weighted radiomic scores based on pre-treatment enhanced CT images can predict tumor responses to immunotherapy in patients with non-small cell lung cancer.

Molecular Signatures in Swine Innate and Adaptive Immune Responses to African Swine Fever Virus Antigens p30/p54/CD2v Expressed Using a Highly Efficient Semliki Forest Virus Replicon System.

African swine fever virus (ASFV) causes a devastating viral hemorrhagic disease in domestic pigs and Eurasian wild boars, posing a foremost threat to the swine industry and pig farming. The development of an effective vaccine is urgently needed, but has been hampered by the lack of an in-depth, mechanistic understanding of the host immune response to ASFV infection and the induction of protective immunity. In this study, we report that immunization of pigs with Semliki Forest Virus (SFV) replicon-based vaccine candidates expressing ASFV p30, p54, and CD2v, as well as their ubiquitin-fused derivatives, elicits T cell differentiation and expansion, promoting specific T cell and humoral immunity. Due to significant variations in the individual non-inbred pigs in response to the vaccination, a personalized analysis was conducted. Using integrated analysis of differentially expressed genes (DEGs), Venn, KEGG and WGCNA, Toll-like receptor, C-type lectin receptor, IL17 receptor, NOD-like receptor and nucleic acid sensor-mediated signaling pathways were demonstrated to be positively correlated to the antigen-stimulated antibody production and inversely correlated to the IFN-γ secreting cell counts in peripheral blood mononuclear cells (PBMCs). An up-regulation of CIQA, CIQB, CIQC, C4BPA, SOSC3, S100A8 and S100A9, and down-regulation of CTLA4, CXCL2, CXCL8, FOS, RGS1, EGR1 and SNAI1 are general in the innate immune response post-the second boost. This study reveals that pattern recognition receptors TLR4, DHX58/DDX58 and ZBP1, and chemokines CXCL2, CXCL8 and CXCL10 may play important roles in regulating this vaccination-stimulated adaptive immune response.

Development and validation of a nonenhanced CT based radiomics model to detect brown adipose tissue.

Purpose: It has been reported that brown adipose tissue (BAT) has a protective effect regarding cardiovascular disease. Positron emission tomography-computed tomography (PET-CT) is the reference method for detecting active BAT; however, it is not feasible to screen for BAT due to the required radionuclides and high-cost. The purpose of this study is to develop and validate a nonenhanced CT based radiomics model to detect BAT and to explore the relationship between CT radiomics derived BAT and cardiovascular calcification. Patients and methods: 146 patients undergoing 18F-FDG PET-CT were retrospectively included from two centers for model development (n = 86) and external validation (n = 60). The data for the model development were randomly divided into a training cohort and an internal validation cohort with a 7:3 ratio, while the external validation data were divided 1:1 into a propensity score matching (PSM) cohort and a randomly sex matched cohort. Radiomics features of BAT and non-BAT depots were extracted from regions of interest (ROI) on nonenhanced CT corresponding to PET studies. Inter-class correlation coefficient (ICC) and Pearson's correlation analysis were performed to select radiomics features with high consistency. Next, least absolute shrinkage and selection operator (LASSO) with linear regression model was used to select radiomics features for model construction. Support vector machine (SVM) was used to develop the model and a radiomics score (RS) was calculated for each depot. The diagnostic performance of the radiomics model was evaluated both on a per-depot and per-patient basis by calculating the area under the receiver operating characteristic curve (AUROC). We further divided patients into BAT-RS group and non-BAT-RS group based on radiomics score and compared their cardiovascular calcification by calculating calcium volume and score. Results: A total of 22 radiomics features were selected for model construction. On a per-depot basis, the AUROCs were 0.87 (95% CI: 0.83-0.9), 0.85 (95% CI: 0.79-0.90), 0.72 (95% CI: 0.67-0.77) and 0.74 (95% CI: 0.69-0.79) for detecting BAT in the training, internal validation, external validation 1 and external validation 2 cohorts, respectively. On a per-patient basis, the radiomics model had high AUROCs of 0.91 (95% CI: 0.84-0.98), 0.77 (95% CI: 0.61-0.92) and 0.85 (95% CI: 0.72-0.98) in the training, external validation 1 and external validation 2 cohorts, respectively. When grouping based on the radiomics model, the BAT-RS group had lower odds of coronary artery calcium (CAC) and thoracic aorta calcium (TAC) compared with the non-BAT-RS group (CAC: 2.8% vs. 20.3%, p = 0.001; TAC: 19.4% vs. 39.2%, p = 0.009). The BAT-RS group had less CAC volume (4.1 ± 4.0 mm3 vs. 147.4 ± 274.3 mm3; p = 0.001), CAC score (2.8 ± 3.0 vs. 169.1 ± 311.5; p = 0.001), TAC volume (301.4 ± 450.2 mm3 vs. 635.3 ± 1100.7 mm3; p = 0.007) and TAC score (496.2 ± 132.6 vs. 749.2 ± 1297.3; p = 0.007) than the non-BAT-RS group. Conclusion: We developed and validated a nonenhanced CT based reliable radiomics model for detecting BAT with PET-CT findings as reference standard. Radiomics signatures from nonenhanced CT can reliably detect BAT and have promising potential to be used in routine clinical settings. Importantly, our study showed that patients with BAT had less cardiovascular calcification.

An automatic fresh rib fracture detection and positioning system using deep learning.

OBJECTIVE: To evaluate the performance and robustness of a deep learning-based automatic fresh rib fracture detection and positioning system (FRF-DPS). METHODS: CT scans of 18,172 participants admitted to eight hospitals from June 2009 to March 2019 were retrospectively collected. Patients were divided into development set (14,241), multicenter internal test set (1612), and external test set (2319). In internal test set, sensitivity, false positives (FPs) and specificity were used to assess fresh rib fracture detection performance at the lesion- and examination-levels. In external test set, the performance of detecting fresh rib fractures by radiologist and FRF-DPS were evaluated at lesion, rib, and examination levels. Additionally, the accuracy of FRF-DPS in rib positioning was investigated by the ground-truth labeling. RESULTS: In multicenter internal test set, FRF-DPS showed excellent performance at the lesion- (sensitivity: 0.933 [95%CI, 0.916-0.949], FPs: 0.50 [95%CI, 0.397-0.583]) and examination-level. In external test set, the sensitivity and FPs at the lesion-level of FRF-DPS (0.909 [95%CI, 0.883-0.926], p < 0.001; 0.379 [95%CI, 0.303-0.422], p = 0.001) were better than the radiologist (0.789 [95%CI, 0.766-0.807]; 0.496 [95%CI, 0.383-0.571]), so were the rib- and patient-levels. In subgroup analysis of CT parameters, FRF-DPS were robust (0.894-0.927). Finally, FRF-DPS(0.997 [95%CI, 0.992-1.000], p < 0.001) is more accurate than radiologist (0.981 [95%CI, 0.969-0.996]) in rib positioning and takes 20 times less time. CONCLUSION: FRF-DPS achieved high detection rate of fresh rib fractures with low FP values, and precise positioning of ribs, thus can be used in clinical practice to improve the detection rate and work efficiency. ADVANCES IN KNOWLEDGE: We developed the FRF-DPS system which can detect fresh rib fractures and rib position, and evaluated by a large amount of multicenter data.

A pre-treatment CT-based weighted radiomic approach combined with clinical characteristics to predict durable clinical benefits of immunotherapy in advanced lung cancer.

OBJECTIVES: To develop a pre-treatment CT-based predictive model to anticipate inoperable lung cancer patients' progression-free survival (PFS) to immunotherapy. METHODS: This single-center retrospective study developed and cross-validated a radiomic model in 185 patients and tested it in 48 patients. The binary endpoint is the durable clinical benefit (DCB, PFS ≥ 6 months) and non-DCB (NDCB, PFS < 6 months). Radiomic features were extracted from multiple intrapulmonary lesions and weighted by an attention-based multiple-instance learning model. Aggregated features were then selected through L2-regularized ridge regression. Five machine-learning classifiers were conducted to build predictive models using radiomic and clinical features alone and then together. Lastly, the predictive value of the model with the best performance was validated by Kaplan-Meier survival analysis. RESULTS: The predictive models based on the weighted radiomic approach showed superior performance across all classifiers (AUCs: 0.75-0.82) compared with the largest lesion approach (AUCs: 0.70-0.78) and the average sum approach (AUCs: 0.64-0.80). Among them, the logistic regression model yielded the most balanced performance (AUC = 0.87 [95%CI 0.84-0.89], 0.75 [0.68-0.82], 0.80 [0.68-0.92] in the training, validation, and test cohort respectively). The addition of five clinical characteristics significantly enhanced the performance of radiomic-only model (train: AUC 0.91 [0.89-0.93], p = .042; validation: AUC 0.86 [0.80-0.91], p = .011; test: AUC 0.86 [0.76-0.96], p = .026). Kaplan-Meier analysis of the radiomic-based predictive models showed a clear stratification between classifier-predicted DCB versus NDCB for PFS (HR = 2.40-2.95, p < 0.05). CONCLUSIONS: The adoption of weighted radiomic features from multiple intrapulmonary lesions has the potential to predict long-term PFS benefits for patients who are candidates for PD-1/PD-L1 immunotherapies. KEY POINTS: • Weighted radiomic-based model derived from multiple intrapulmonary lesions on pre-treatment CT images has the potential to predict durable clinical benefits of immunotherapy in lung cancer. • Early line immunotherapy is associated with longer progression-free survival in advanced lung cancer.

Total nodule number as an independent prognostic factor in resected stage III non-small cell lung cancer: a deep learning-powered study.

Background: Almost every patient with lung cancer has multiple pulmonary nodules; however, the significance of nodule multiplicity in locally advanced non-small cell lung cancer (NSCLC) remains unclear. Methods: We identified patients who had undergone surgical resection for stage I-III NSCLC at the Peking University People's Hospital from 2005 to 2018 for whom preoperative chest computed tomography (CT) scans were available. Deep learning-based artificial intelligence (AI) algorithms using convolutional neural networks (CNN) were applied to detect and classify pulmonary nodules (PNs). Maximally selected log-rank statistics were used to determine the optimal cutoff value of the total nodule number (TNN) for predicting survival. Results: A total of 33,410 PNs were detected by AI among the 2,126 participants. The median TNN detected per person was 12 [interquartile range (IQR) 7-20]. It was revealed that AI-detected TNN (analyzed as a continuous variable) was an independent prognostic factor for both recurrence-free survival (RFS) [hazard ratio (HR) 1.012, 95% confidence interval (CI): 1.002 to 1.022, P=0.021] and overall survival (OS) (HR 1.013, 95% CI: 1.002 to 1.025, P=0.021) in multivariate analyses of the stage III cohort. In contrast, AI-detected TNN was not significantly associated with survival in the stage I and II cohorts. In a survival tree analysis, rather than using traditional IIIA and IIIB classifications, the model grouped cases according to AI-detected TNN (lower vs. higher: log-rank P<0.001), which led to a more effective determination of survival rates in the stage III cohort. Conclusions: The AI-detected TNN is significantly associated with survival rates in patients with surgically resected stage III NSCLC. A lower TNN detected on preoperative CT scans indicates a better prognosis for patients who have undergone complete surgical resection.

An original deep learning model using limited data for COVID-19 discrimination: A multicenter study.

OBJECTIVES: Artificial intelligence (AI) has been proved to be a highly efficient tool for COVID-19 diagnosis, but the large data size and heavy label force required for algorithm development and the poor generalizability of AI algorithms, to some extent, limit the application of AI technology in clinical practice. The aim of this study is to develop an AI algorithm with high robustness using limited chest CT data for COVID-19 discrimination. METHODS: A three dimensional algorithm that combined multi-instance learning with the LSTM architecture (3DMTM) was developed for differentiating COVID-19 from community acquired pneumonia (CAP) while logistic regression (LR), k-nearest neighbor (KNN), support vector machine (SVM), and a three dimensional convolutional neural network set for comparison. Totally, 515 patients with or without COVID-19 between December 2019 and March 2020 from five different hospitals were recruited and divided into relatively large (150 COVID-19 and 183 CAP cases) and relatively small datasets (17 COVID-19 and 35 CAP cases) for either training or validation and another independent dataset (37 COVID-19 and 93 CAP cases) for external test. Area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision, accuracy, F1 score, and G-mean were utilized for performance evaluation. RESULTS: In the external test cohort, the relatively large data-based 3DMTM-LD achieved an AUC of 0.956 (95% confidence interval, 95% CI, 0.929∼0.982) with 86.2% and 98.0% for its sensitivity and specificity. 3DMTM-SD got an AUC of 0.937 (95% CI, 0.909∼0.965), while the AUC of 3DCM-SD decreased dramatically to 0.714 (95% CI, 0.649∼0.780) with training data reduction. KNN-MMSD, LR-MMSD, SVM-MMSD, and 3DCM-MMSD benefited significantly from the inclusion of clinical information while models trained with relatively large dataset got slight performance improvement in COVID-19 discrimination. 3DMTM, trained with either CT or multi-modal data, presented comparably excellent performance in COVID-19 discrimination. CONCLUSIONS: The 3DMTM algorithm presented excellent robustness for COVID-19 discrimination with limited CT data. 3DMTM based on CT data performed comparably in COVID-19 discrimination with that trained with multi-modal information. Clinical information could improve the performance of KNN, LR, SVM, and 3DCM in COVID-19 discrimination, especially in the scenario with limited data for training.

Detection and staging of chronic obstructive pulmonary disease using a computed tomography-based weakly supervised deep learning approach.

OBJECTIVES: Chronic obstructive pulmonary disease (COPD) is underdiagnosed globally. The present study aimed to develop weakly supervised deep learning (DL) models that utilize computed tomography (CT) image data for the automated detection and staging of spirometry-defined COPD. METHODS: A large, highly heterogeneous dataset was established, consisting of 1393 participants retrospectively recruited from outpatient, inpatient, and physical examination center settings of four large public hospitals in China. All participants underwent both inspiratory chest CT scans and pulmonary function tests. CT images, spirometry data, demographic information, and clinical information of each participant were collected. An attention-based multi-instance learning (MIL) model for COPD detection was trained using CT scans from 837 participants. External validation of the COPD detection was performed with 620 low-dose CT (LDCT) scans acquired from the National Lung Screening Trial (NLST) cohort. A multi-channel 3D residual network was further developed to categorize GOLD stages among confirmed COPD patients. RESULTS: The attention-based MIL model used for COPD detection achieved an area under the receiver operating characteristic curve (AUC) of 0.934 (95% CI: 0.903, 0.961) on the internal test set and 0.866 (95% CI: 0.805, 0.928) on the LDCT subset acquired from the NLST. The multi-channel 3D residual network was able to correctly grade 76.4% of COPD patients in the test set (423/553) using the GOLD scale. CONCLUSIONS: The proposed chest CT-DL approach can automatically identify spirometry-defined COPD and categorize patients according to the GOLD scale. As such, this approach may be an effective case-finding tool for COPD diagnosis and staging. KEY POINTS: • Chronic obstructive pulmonary disease is underdiagnosed globally, particularly in developing countries. • The proposed chest computed tomography (CT)-based deep learning (DL) approaches could accurately identify spirometry-defined COPD and categorize patients according to the GOLD scale. • The chest CT-DL approach may be an alternative case-finding tool for COPD identification and evaluation.

Diagnosing autism spectrum disorder in children using conventional MRI and apparent diffusion coefficient based deep learning algorithms.

OBJECTIVE: To develop and validate deep learning (DL) methods for diagnosing autism spectrum disorder (ASD) based on conventional MRI (cMRI) and apparent diffusion coefficient (ADC) images. METHODS: A total of 151 ASD children and 151 age-matched typically developing (TD) controls were included in this study. The data from these subjects were assigned to training and validation datasets. An additional 20 ASD children and 25 TD controls were acquired, whose data were utilized in an independent test set. All subjects underwent cMRI and diffusion-weighted imaging examination of the brain. We developed a series of DL models to separate ASD from TD based on the cMRI and ADC data. The seven models used include five single-sequence models (SSMs), one dominant-sequence model (DSM), and one all-sequence model (ASM). To enhance the feature detection of the models, we embed an attention mechanism module. RESULTS: The highest AUC (0.824 ~ 0.850) was achieved when applying the SSM based on either FLAIR or ADC to the validation and independent test sets. A DSM using the combination of FLAIR and ADC showed an improved AUC in the validation (0.873) and independent test sets (0.876). The ASM also showed better diagnostic value in the validation (AUC = 0.838) and independent test sets (AUC = 0.836) compared to the SSMs. Among the models with attention mechanism, the DSM achieved the highest diagnostic performance with an AUC, accuracy, sensitivity, and specificity of 0.898, 84.4%, 85.0%, and 84.0% respectively. CONCLUSIONS: This study established the potential of DL models to distinguish ASD cases from TD controls based on cMRI and ADC images. KEY POINTS: • Deep learning models based on conventional MRI and ADC can be used to diagnose ASD. • The model (DSM) based on the FLAIR and ADC sequence achieved the best diagnostic performance with an AUC of 0.836 in the independent test sets. • The attention mechanism further improved the diagnostic performance of the models.

Deep Learning-Enabled Identification of Autoimmune Encephalitis on 3D Multi-Sequence MRI.

BACKGROUND: Autoimmune encephalitis (AE) is a noninfectious emergency with severe clinical attacks. It is difficult for the earlier diagnosis of acute AE due to the lack of antibody detection resources. PURPOSE: To construct a deep learning (DL) algorithm using multi-sequence magnetic resonance imaging (MRI) for the identification of acute AE. STUDY TYPE: Retrospective. POPULATION: One hundred and sixty AE patients (90 women; median age 36), 177 herpes simplex virus encephalitis (HSVE) (89 women; median age 39), and 184 healthy controls (HC) (95 women; median age 39) were included. Fifty-two patients from another site were enrolled for external validation. FIELD STRENGTH/SEQUENCE: 3.0 T; fast spin-echo (T1 WI, T2 WI, fluid attenuated inversion recovery imaging) and spin-echo echo-planar diffusion weighted imaging. ASSESSMENT: Five DL models based on individual or combined four MRI sequences to classify the datasets as AE, HSVE, or HC. Reader experiment was further carried out by radiologists. STATISTICAL TESTS: The discriminative performance of different models was assessed using the area under the receiver operating characteristic curve (AUC). The optimal threshold cut-off was identified when sensitivity and specificity were maximized (sensitivity + specificity - 1) in the validation set. Classification performance using confusion matrices was reported to evaluate the diagnostic value of the models and the radiologists' assessments before being assessed by the paired t-test (P < 0.05 was considered significant). RESULTS: In the internal test set, the fusion model achieved the significantly greatest diagnostic performance than single-sequence DL models with AUCs of 0.828, 0.884, and 0.899 for AE, HSVE, and HC, respectively. The model demonstrated a consistently high performance in the external validation set with AUCs of 0.831 (AE), 0.882 (HSVE), and 0.892 (HC). The fusion model also demonstrated significantly higher performance than all radiologists in identifying AE (accuracy between the fuse model vs. average radiologist: 83% vs. 72%). DATA CONCLUSION: The proposed DL algorithm derived from multi-sequence MRI provided desirable identification and classification of acute AE. LEVEL OF EVIDENCE: 3 TECHNICAL EFFICACY: Stage 2.

Development and validation of MRI-based deep learning models for prediction of microsatellite instability in rectal cancer.

BACKGROUND: Microsatellite instability (MSI) predetermines responses to adjuvant 5-fluorouracil and immunotherapy in rectal cancer and serves as a prognostic biomarker for clinical outcomes. Our objective was to develop and validate a deep learning model that could preoperatively predict the MSI status of rectal cancer based on magnetic resonance images. METHODS: This single-center retrospective study included 491 rectal cancer patients with pathologically proven microsatellite status. Patients were randomly divided into the training/validation cohort (n = 395) and the testing cohort (n = 96). A clinical model using logistic regression was constructed to discriminate MSI status using only clinical factors. Based on a modified MobileNetV2 architecture, deep learning models were tested for the predictive ability of MSI status from magnetic resonance images, with or without integrating clinical factors. RESULTS: The clinical model correctly classified 37.5% of MSI status in the testing cohort, with an AUC value of 0.573 (95% confidence interval [CI], 0.468 ~ 0.674). The pure imaging-based model and the combined model correctly classified 75.0% and 85.4% of MSI status in the testing cohort, with AUC values of 0.820 (95% CI, 0.718 ~ 0.884) and 0.868 (95% CI, 0.784 ~ 0.929), respectively. Both deep learning models performed better than the clinical model (p < 0.05). There was no statistically significant difference between the deep learning models with or without integrating clinical factors. CONCLUSIONS: Deep learning based on high-resolution T2-weighted magnetic resonance images showed a good predictive performance for MSI status in rectal cancer patients. The proposed model may help to identify patients who would benefit from chemotherapy or immunotherapy and determine individualized therapeutic strategies for these patients.

A deep learning integrated radiomics model for identification of coronavirus disease 2019 using computed tomography.

Since its first outbreak, Coronavirus Disease 2019 (COVID-19) has been rapidly spreading worldwide and caused a global pandemic. Rapid and early detection is essential to contain COVID-19. Here, we first developed a deep learning (DL) integrated radiomics model for end-to-end identification of COVID-19 using CT scans and then validated its clinical feasibility. We retrospectively collected CT images of 386 patients (129 with COVID-19 and 257 with other community-acquired pneumonia) from three medical centers to train and externally validate the developed models. A pre-trained DL algorithm was utilized to automatically segment infected lesions (ROIs) on CT images which were used for feature extraction. Five feature selection methods and four machine learning algorithms were utilized to develop radiomics models. Trained with features selected by L1 regularized logistic regression, classifier multi-layer perceptron (MLP) demonstrated the optimal performance with AUC of 0.922 (95% CI 0.856-0.988) and 0.959 (95% CI 0.910-1.000), the same sensitivity of 0.879, and specificity of 0.900 and 0.887 on internal and external testing datasets, which was equivalent to the senior radiologist in a reader study. Additionally, diagnostic time of DL-MLP was more efficient than radiologists (38 s vs 5.15 min). With an adequate performance for identifying COVID-19, DL-MLP may help in screening of suspected cases.

A deep learning model integrating mammography and clinical factors facilitates the malignancy prediction of BI-RADS 4 microcalcifications in breast cancer screening.

OBJECTIVES: To investigate the value of full-field digital mammography-based deep learning (DL) in predicting malignancy of Breast Imaging Reporting and Data System (BI-RADS) 4 microcalcifications. METHODS: A total of 384 patients with 414 pathologically confirmed microcalcifications (221 malignant and 193 benign) were randomly allocated into the training, validation, and testing datasets (272/71/71 lesions) in this retrospective study. A combined DL model was developed incorporating mammography and clinical variables. Model performance was evaluated by using areas under the receiver operating characteristic curve (AUC) and compared with the clinical model, stand-alone DL image model, and BI-RADS approach. The predictive performance for malignancy was also compared between the combined model and human readers (2 juniors and 2 seniors). RESULTS: The combined DL model demonstrated favorable AUC, sensitivity, and specificity of 0.910, 85.3%, and 91.9% in predicting BI-RADS 4 malignant microcalcifications in the testing dataset, which outperformed the clinical model, DL image model, and BI-RADS with AUCs of 0.799, 0.841, and 0.804, respectively. The combined model achieved non-inferior performance as senior radiologists (p = 0.860, p = 0.800) and outperformed junior radiologists (p = 0.155, p = 0.029). The diagnostic performance of two junior radiologists was improved after artificial intelligence assistance with AUCs increased to 0.854 and 0.901 from 0.816 (p = 0.556) and 0.773 (p = 0.046), while the interobserver agreement was improved with a kappa value increased to 0.843 from 0.331. CONCLUSIONS: The combined deep learning model can improve the malignancy prediction of BI-RADS 4 microcalcifications in screening mammography and assist junior radiologists to achieve better performance, which can facilitate clinical decision-making. KEY POINTS: • The combined deep learning model demonstrated high diagnostic power, sensitivity, and specificity for predicting malignant BI-RADS 4 mammographic microcalcifications. • The combined model achieved similar performance with senior breast radiologists, while it outperformed junior breast radiologists. • Deep learning could improve the diagnostic performance of junior radiologists and facilitate clinical decision-making.