Deep learning-based prognostication in idiopathic pulmonary fibrosis using chest radiographs.

OBJECTIVES: To develop and validate a deep learning-based prognostic model in patients with idiopathic pulmonary fibrosis (IPF) using chest radiographs. METHODS: To develop a deep learning-based prognostic model using chest radiographs (DLPM), the patients diagnosed with IPF during 2011-2021 were retrospectively collected and were divided into training (n = 1007), validation (n = 117), and internal test (n = 187) datasets. Up to 10 consecutive radiographs were included for each patient. For external testing, three cohorts from independent institutions were collected (n = 152, 141, and 207). The discrimination performance of DLPM was evaluated using areas under the time-dependent receiver operating characteristic curves (TD-AUCs) for 3-year survival and compared with that of forced vital capacity (FVC). Multivariable Cox regression was performed to investigate whether the DLPM was an independent prognostic factor from FVC. We devised a modified gender-age-physiology (GAP) index (GAP-CR), by replacing DLCO with DLPM. RESULTS: DLPM showed similar-to-higher performance at predicting 3-year survival than FVC in three external test cohorts (TD-AUC: 0.83 [95% CI: 0.76-0.90] vs. 0.68 [0.59-0.77], p < 0.001; 0.76 [0.68-0.85] vs. 0.70 [0.60-0.80], p = 0.21; 0.79 [0.72-0.86] vs. 0.76 [0.69-0.83], p = 0.41). DLPM worked as an independent prognostic factor from FVC in all three cohorts (ps < 0.001). The GAP-CR index showed a higher 3-year TD-AUC than the original GAP index in two of the three external test cohorts (TD-AUC: 0.85 [0.80-0.91] vs. 0.79 [0.72-0.86], p = 0.02; 0.72 [0.64-0.80] vs. 0.69 [0.61-0.78], p = 0.56; 0.76 [0.69-0.83] vs. 0.68 [0.60-0.76], p = 0.01). CONCLUSIONS: A deep learning model successfully predicted survival in patients with IPF from chest radiographs, comparable to and independent of FVC. CLINICAL RELEVANCE STATEMENT: Deep learning-based prognostication from chest radiographs offers comparable-to-higher prognostic performance than forced vital capacity. KEY POINTS: • A deep learning-based prognostic model for idiopathic pulmonary fibrosis was developed using 6063 radiographs. • The prognostic performance of the model was comparable-to-higher than forced vital capacity, and was independent from FVC in all three external test cohorts. • A modified gender-age-physiology index replacing diffusing capacity for carbon monoxide with the deep learning model showed higher performance than the original index in two external test cohorts.

Diagnostic performance and prognostic value of CT-defined visceral pleural invasion in early-stage lung adenocarcinomas.

OBJECTIVES: To analyze the diagnostic performance and prognostic value of CT-defined visceral pleural invasion (CT-VPI) in early-stage lung adenocarcinomas. METHODS: Among patients with clinical stage I lung adenocarcinomas, half of patients were randomly selected for a diagnostic study, in which five thoracic radiologists determined the presence of CT-VPI. Probabilities for CT-VPI were obtained using deep learning (DL). Areas under the receiver operating characteristic curve (AUCs) and binary diagnostic measures were calculated and compared. Inter-rater agreement was assessed. For all patients, the prognostic value of CT-VPI by two radiologists and DL (using high-sensitivity and high-specificity cutoffs) was investigated using Cox regression. RESULTS: In 681 patients (median age, 65 years [interquartile range, 58-71]; 382 women), pathologic VPI was positive in 130 patients. For the diagnostic study (n = 339), the pooled AUC of five radiologists was similar to that of DL (0.78 vs. 0.79; p = 0.76). The binary diagnostic performance of radiologists was variable (sensitivity, 45.3-71.9%; specificity, 71.6-88.7%). Inter-rater agreement was moderate (weighted Fleiss κ, 0.51; 95%CI: 0.43-0.55). For overall survival (n = 680), CT-VPI by radiologists (adjusted hazard ratio [HR], 1.27 and 0.99; 95%CI: 0.84-1.92 and 0.63-1.56; p = 0.26 and 0.97) or DL (HR, 1.44 and 1.06; 95%CI: 0.86-2.42 and 0.67-1.68; p = 0.17 and 0.80) was not prognostic. CT-VPI by an attending radiologist was prognostic only in radiologically solid tumors (HR, 1.82; 95%CI: 1.07-3.07; p = 0.03). CONCLUSION: The diagnostic performance and prognostic value of CT-VPI are limited in clinical stage I lung adenocarcinomas. This feature may be applied for radiologically solid tumors, but substantial reader variability should be overcome. CLINICAL RELEVANCE STATEMENT: Although the diagnostic performance and prognostic value of CT-VPI are limited in clinical stage I lung adenocarcinomas, this parameter may be applied for radiologically solid tumors with appropriate caution regarding inter-reader variability. KEY POINTS: • Use of CT-defined visceral pleural invasion in clinical staging should be cautious, because prognostic value of CT-defined visceral pleural invasion remains unexplored. • Diagnostic performance and prognostic value of CT-defined visceral pleural invasion varied among radiologists and deep learning. • Role of CT-defined visceral pleural invasion in clinical staging may be limited to radiologically solid tumors.

AI Improves Nodule Detection on Chest Radiographs in a Health Screening Population: A Randomized Controlled Trial.

Background The impact of artificial intelligence (AI)-based computer-aided detection (CAD) software has not been prospectively explored in real-world populations. Purpose To investigate whether commercial AI-based CAD software could improve the detection rate of actionable lung nodules on chest radiographs in participants undergoing health checkups. Materials and Methods In this single-center, pragmatic, open-label randomized controlled trial, participants who underwent chest radiography between July 2020 and December 2021 in a health screening center were enrolled and randomized into intervention (AI group) and control (non-AI group) arms. One of three designated radiologists with 13-36 years of experience interpreted each radiograph, referring to the AI-based CAD results for the AI group. The primary outcome was the detection rate, that is, the number of true-positive radiographs divided by the total number of radiographs, of actionable lung nodules confirmed on CT scans obtained within 3 months. Actionable nodules were defined as solid nodules larger than 8 mm or subsolid nodules with a solid portion larger than 6 mm (Lung Imaging Reporting and Data System, or Lung-RADS, category 4). Secondary outcomes included the positive-report rate, sensitivity, false-referral rate, and malignant lung nodule detection rate. Clinical outcomes were compared between the two groups using univariable logistic regression analyses. Results A total of 10 476 participants (median age, 59 years [IQR, 50-66 years]; 5121 men) were randomized to an AI group (n = 5238) or non-AI group (n = 5238). The trial met the predefined primary outcome, demonstrating an improved detection rate of actionable nodules in the AI group compared with the non-AI group (0.59% [31 of 5238 participants] vs 0.25% [13 of 5238 participants], respectively; odds ratio, 2.4; 95% CI: 1.3, 4.7; P = .008). The detection rate for malignant lung nodules was higher in the AI group compared with the non-AI group (0.15% [eight of 5238 participants] vs 0.0% [0 of 5238 participants], respectively; P = .008). The AI and non-AI groups showed similar false-referral rates (45.9% [56 of 122 participants] vs 56.0% [56 of 100 participants], respectively; P = .14) and positive-report rates (2.3% [122 of 5238 participants] vs 1.9% [100 of 5238 participants]; P = .14). Conclusion In health checkup participants, artificial intelligence-based software improved the detection of actionable lung nodules on chest radiographs. © RSNA, 2023 Supplemental material is available for this article. See also the editorial by Auffermann in this isssue.

Prognostic value of deep learning-based fibrosis quantification on chest CT in idiopathic pulmonary fibrosis.

OBJECTIVE: To investigate the prognostic value of deep learning (DL)-driven CT fibrosis quantification in idiopathic pulmonary fibrosis (IPF). METHODS: Patients diagnosed with IPF who underwent nonenhanced chest CT and spirometry between 2005 and 2009 were retrospectively collected. Proportions of normal (CT-Norm%) and fibrotic lung volume (CT-Fib%) were calculated on CT using the DL software. The correlations of CT-Norm% and CT-Fib% with forced vital capacity (FVC) and diffusion capacity of carbon monoxide (DLCO) were evaluated. The multivariable-adjusted hazard ratios (HRs) of CT-Norm% and CT-Fib% for overall survival were calculated with clinical and physiologic variables as covariates using Cox regression. The feasibility of substituting CT-Norm% for DLCO in the GAP index was investigated using time-dependent areas under the receiver operating characteristic curve (TD-AUCs) at 3 years. RESULTS: In total, 161 patients (median age [IQR], 68 [62-73] years; 104 men) were evaluated. CT-Norm% and CT-Fib% showed significant correlations with FVC (Pearson's r, 0.40 for CT-Norm% and - 0.37 for CT-Fib%; both p < 0.001) and DLCO (0.52 for CT-Norm% and - 0.46 for CT-Fib%; both p < 0.001). On multivariable Cox regression, both CT-Norm% and CT-Fib% were independent prognostic factors when adjusted to age, sex, smoking status, comorbid chronic diseases, FVC, and DLCO (HRs, 0.98 [95% CI 0.97-0.99; p < 0.001] for CT-Norm% at 3 years and 1.03 [1.01-1.05; p = 0.01] for CT-Fib%). Substituting CT-Norm% for DLCO showed comparable discrimination to the original GAP index (TD-AUC, 0.82 [0.78-0.85] vs. 0.82 [0.79-0.86]; p = 0.75). CONCLUSION: CT-Norm% and CT-Fib% calculated using chest CT-based deep learning software were independent prognostic factors for overall survival in IPF. KEY POINTS: • Normal and fibrotic lung volume proportions were automatically calculated using commercial deep learning software from chest CT taken from 161 patients diagnosed with idiopathic pulmonary fibrosis. • CT-quantified volumetric parameters from commercial deep learning software were correlated with forced vital capacity (Pearson's r, 0.40 for normal and - 0.37 for fibrotic lung volume proportions) and diffusion capacity of carbon monoxide (Pearson's r, 0.52 and - 0.46, respectively). • Normal and fibrotic lung volume proportions (hazard ratios, 0.98 and 1.04; both p < 0.001) independently predicted overall survival when adjusted for clinical and physiologic variables.

75% radiation dose reduction using deep learning reconstruction on low-dose chest CT.

OBJECTIVE: Few studies have explored the clinical feasibility of using deep-learning reconstruction to reduce the radiation dose of CT. We aimed to compare the image quality and lung nodule detectability between chest CT using a quarter of the low dose (QLD) reconstructed with vendor-agnostic deep-learning image reconstruction (DLIR) and conventional low-dose (LD) CT reconstructed with iterative reconstruction (IR). MATERIALS AND METHODS: We retrospectively collected 100 patients (median age, 61 years [IQR, 53-70 years]) who received LDCT using a dual-source scanner, where total radiation was split into a 1:3 ratio. QLD CT was generated using a quarter dose and reconstructed with DLIR (QLD-DLIR), while LDCT images were generated using a full dose and reconstructed with IR (LD-IR). Three thoracic radiologists reviewed subjective noise, spatial resolution, and overall image quality, and image noise was measured in five areas. The radiologists were also asked to detect all Lung-RADS category 3 or 4 nodules, and their performance was evaluated using area under the jackknife free-response receiver operating characteristic curve (AUFROC). RESULTS: The median effective dose was 0.16 (IQR, 0.14-0.18) mSv for QLD CT and 0.65 (IQR, 0.57-0.71) mSv for LDCT. The radiologists' evaluations showed no significant differences in subjective noise (QLD-DLIR vs. LD-IR, lung-window setting; 3.23 ± 0.19 vs. 3.27 ± 0.22; P = .11), spatial resolution (3.14 ± 0.28 vs. 3.16 ± 0.27; P = .12), and overall image quality (3.14 ± 0.21 vs. 3.17 ± 0.17; P = .15). QLD-DLIR demonstrated lower measured noise than LD-IR in most areas (P < .001 for all). No significant difference was found between QLD-DLIR and LD-IR for the sensitivity (76.4% vs. 72.2%; P = .35) or the AUFROCs (0.77 vs. 0.78; P = .68) in detecting Lung-RADS category 3 or 4 nodules. Under a noninferiority limit of -0.1, QLD-DLIR showed noninferior detection performance (95% CI for AUFROC difference, -0.04 to 0.06). CONCLUSION: QLD-DLIR images showed comparable image quality and noninferior nodule detectability relative to LD-IR images.

Histopathologic Basis for a Chest CT Deep Learning Survival Prediction Model in Patients with Lung Adenocarcinoma.

Background A preoperative CT-based deep learning (DL) prediction model was proposed to estimate disease-free survival in patients with resected lung adenocarcinoma. However, the black-box nature of DL hinders interpretation of its results. Purpose To provide histopathologic evidence underpinning the DL survival prediction model and to demonstrate the feasibility of the model in identifying patients with histopathologic risk factors through unsupervised clustering and a series of regression analyses. Materials and Methods For this retrospective study, data from patients who underwent curative resection for lung adenocarcinoma without neoadjuvant therapy from January 2016 to September 2020 were collected from a tertiary care center. Seven histopathologic risk factors for the resected adenocarcinoma were documented: the aggressive adenocarcinoma subtype (cribriform, morular, solid, or micropapillary-predominant subtype); mediastinal nodal metastasis (pN2); presence of lymphatic, venous, and perineural invasion; visceral pleural invasion (VPI); and EGFR mutation status. Unsupervised clustering using 80 DL model-driven CT features was performed, and associations between the patient clusters and the histopathologic features were analyzed. Multivariable regression analyses were performed to investigate the added value of the DL model output to the semantic CT features (clinical T category and radiologic nodule type [ie, solid or subsolid]) for histopathologic associations. Results A total of 1667 patients (median age, 64 years [IQR, 57-71 years]; 975 women) were evaluated. Unsupervised patient clusters 3 and 4 were associated with all histopathologic risk factors (P < .01) except for EGFR mutation status (P = .30 for cluster 3). After multivariable adjustment, model output was associated with the aggressive adenocarcinoma subtype (odds ratio [OR], 1.03; 95% CI: 1.002, 1.05; P = .03), venous invasion (OR, 1.03; 95% CI: 1.004, 1.06; P = .02), and VPI (OR, 1.08; 95% CI: 1.06, 1.10; P < .001), independently of the semantic CT features. Conclusion The deep learning model extracted CT imaging surrogates for the histopathologic profiles of lung adenocarcinoma. © RSNA, 2022 Online supplemental material is available for this article. See also the editorial by Yanagawa in this issue.

No Prognostic Impact of Staging Brain MRI in Patients with Stage IA Non-Small Cell Lung Cancer.

Background Although various guidelines discourage performing brain MRI for staging purposes in asymptomatic patients with clinical stage IA non-small cell lung cancer (NSCLC), evidence regarding their postoperative survival is lacking. Purpose To investigate the survival benefit of performing brain MRI in asymptomatic patients with early-stage NSCLC. Materials and Methods Patients who underwent curative resection between February 2009 and March 2016 for clinical TNM stage T1N0M0 NSCLC were retrospectively included. Patient survival and development of brain metastasis during postoperative surveillance were documented. The cumulative survival rate and incidence of brain metastasis were compared between patients who underwent surgery with or without staging brain MRI by using Cox regression and a Fine-Gray subdistribution hazard model, respectively, for multivariable adjustment. Propensity score matching and inverse probability of treatment weighting were applied for confounder adjustment. Results A total of 628 patients (mean age, 64 years ± 10 [SD]; 319 men) were included, of whom 53% (331 of 628) underwent staging brain MRI. In the multivariable analyses, brain MRI did not show prognostic benefits for brain metastasis-free survival (hazard ratio [HR], 1.06; 95% CI: 0.69, 1.63; P = .79), time to brain metastasis (HR, 1.60; 95% CI: 0.70, 3.94; P = .29), and overall survival (HR, 0.86; 95% CI, 0.54, 1.37; P = .54). Consistent results were obtained after propensity score matching (brain metastasis-free survival [HR, 0.97; 95% CI: 0.60, 1.57; P = .91], time to brain metastasis [HR, 1.29; 95% CI: 0.50, 3.33; P = .60], and overall survival [HR, 0.89; 95% CI: 0.53, 1.51; P = .67]) and inverse probability of treatment weighting. Conclusion No difference was observed between asymptomatic patients with clinical stage IA non-small cell lung cancer who underwent staging brain MRI and those who did not in terms of brain metastasis-free survival, time to brain metastasis, and overall survival. © RSNA, 2022 Online supplemental material is available for this article. See also the editorial by Bizzi and Pascuzzo in this issue.

Deep Learning Prediction of Survival in Patients with Chronic Obstructive Pulmonary Disease Using Chest Radiographs.

Background Preexisting indexes for predicting the prognosis of chronic obstructive pulmonary disease (COPD) do not use radiologic information and are impractical because they involve complex history assessments or exercise tests. Purpose To develop and to validate a deep learning-based survival prediction model in patients with COPD (DLSP) using chest radiographs, in addition to other clinical factors. Materials and Methods In this retrospective study, data from patients with COPD who underwent postbronchodilator spirometry and chest radiography from 2011-2015 were collected and split into training (n = 3475), validation (n = 435), and internal test (n = 315) data sets. The algorithm for predicting survival from chest radiographs was trained (hereafter, DLSPCXR), and then age, body mass index, and forced expiratory volume in 1 second (FEV1) were integrated within the model (hereafter, DLSPinteg). For external test, three independent cohorts were collected (n = 394, 416, and 337). The discrimination performance of DLSPCXR was evaluated by using time-dependent area under the receiver operating characteristic curves (TD AUCs) at 5-year survival. Goodness of fit was assessed by using the Hosmer-Lemeshow test. Using one external test data set, DLSPinteg was compared with four COPD-specific clinical indexes: BODE, ADO, COPD Assessment Test (CAT), and St George's Respiratory Questionnaire (SGRQ). Results DLSPCXR had a higher performance at predicting 5-year survival than FEV1 in two of the three external test cohorts (TD AUC: 0.73 vs 0.63 [P = .004]; 0.67 vs 0.60 [P = .01]; 0.76 vs 0.77 [P = .91]). DLSPCXR demonstrated good calibration in all cohorts. The DLSPinteg model showed no differences in TD AUC compared with BODE (0.87 vs 0.80; P = .34), ADO (0.86 vs 0.89; P = .51), and SGRQ (0.86 vs 0.70; P = .09), and showed higher TD AUC than CAT (0.93 vs 0.55; P < .001). Conclusion A deep learning model using chest radiographs was capable of predicting survival in patients with chronic obstructive pulmonary disease. © RSNA, 2022 Online supplemental material is available for this article.

Value of a deep learning-based algorithm for detecting Lung-RADS category 4 nodules on chest radiographs in a health checkup population: estimation of the sample size for a randomized controlled trial.

OBJECTIVE: To explore the value of a deep learning-based algorithm in detecting Lung CT Screening Reporting and Data System category 4 nodules on chest radiographs from an asymptomatic health checkup population. METHODS: Data from an annual retrospective cohort of individuals who underwent chest radiographs for health checkup purposes and chest CT scanning within 3 months were collected. Among 3073 individuals, 118 with category 4 nodules on CT were selected. A reader performance test was performed using those 118 radiographs and randomly selected 51 individuals without any nodules. Four radiologists independently evaluated the radiographs without and with the results of the algorithm; and sensitivities/specificities were compared. The sample size needed to confirm the difference in detection rates was calculated, i.e., the number of true-positive radiographs divided by the total number of radiographs. RESULTS: The sensitivity of the radiologists substantially increased aided by the algorithm (38.8% [183/472] to 45.1% [213/472]; p < .001) without significant change in specificity (94.1% [192/204] vs. 92.2% [188/204]; p = .22). Pooled radiologists detected more nodules with the algorithm (32.0% [156/488] vs. 38.9% [190/488]; p < .001), without alteration of false-positive rates (0.09 [62/676], both). Pooled detection rates for the annual cohort were 1.49% (183/12,292) and 1.73% (213/12,292) without and with the algorithm, respectively. A sample size of 41,776 in each arm would be required to demonstrate significant detection rate difference with < 5% type I error and > 80% power. CONCLUSION: Although readers substantially increased sensitivity in detecting nodules on chest radiographs from a health checkup population aided by the algorithm, detection rate difference was only 0.24%, requiring a sample size >80,000 for a randomized controlled trial. KEY POINTS: • Aided by a deep learning algorithm, pooled radiologists improved their sensitivity in detecting Lung-RADS category 4 nodules on chest radiographs from a health checkup population (38.8% [183/472] to 45.1% [213/472]; p < .001), without increasing false-positive rate. • The prevalence of the Lung-RADS category 4 nodules was 3.8% (118/3073) on the population, resulting in only 0.24% increase of the detection rate for the radiologists with assistance of the algorithm. • To confirm the significant detection rate increase by a randomized controlled trial, a sample size of 84,000 would be required.

Artificial intelligence system for identification of false-negative interpretations in chest radiographs.

OBJECTIVES: To investigate the efficacy of an artificial intelligence (AI) system for the identification of false negatives in chest radiographs that were interpreted as normal by radiologists. METHODS: We consecutively collected chest radiographs that were read as normal during 1 month (March 2020) in a single institution. A commercialized AI system was retrospectively applied to these radiographs. Radiographs with abnormal AI results were then re-interpreted by the radiologist who initially read the radiograph ("AI as the advisor" scenario). The reference standards for the true presence of relevant abnormalities in radiographs were defined by majority voting of three thoracic radiologists. The efficacy of the AI system was evaluated by detection yield (proportion of true-positive identification among the entire examination) and false-referral rate (FRR, proportion of false-positive identification among all examinations). Decision curve analyses were performed to evaluate the net benefits of applying the AI system. RESULTS: A total of 4208 radiographs from 3778 patients (M:F = 1542:2236; median age, 56 years) were included. The AI system identified initially overlooked relevant abnormalities with a detection yield and an FRR of 2.4% and 14.0%, respectively. In the "AI as the advisor" scenario, radiologists detected initially overlooked relevant abnormalities with a detection yield and FRR of 1.2% and 0.97%, respectively. In a decision curve analysis, AI as an advisor scenario exhibited a positive net benefit when the cost-to-benefit ratio was below 1:0.8. CONCLUSION: An AI system could identify relevant abnormalities overlooked by radiologists and could enable radiologists to correct their false-negative interpretations by providing feedback to radiologists. KEY POINTS: • In consecutive chest radiographs with normal interpretations, an artificial intelligence system could identify relevant abnormalities that were initially overlooked by radiologists. • The artificial intelligence system could enable radiologists to correct their initial false-negative interpretations by providing feedback to radiologists when overlooked abnormalities were present.

Development and validation of a deep learning algorithm detecting 10 common abnormalities on chest radiographs.

We aimed to develop a deep learning algorithm detecting 10 common abnormalities (DLAD-10) on chest radiographs, and to evaluate its impact in diagnostic accuracy, timeliness of reporting and workflow efficacy.DLAD-10 was trained with 146 717 radiographs from 108 053 patients using a ResNet34-based neural network with lesion-specific channels for 10 common radiological abnormalities (pneumothorax, mediastinal widening, pneumoperitoneum, nodule/mass, consolidation, pleural effusion, linear atelectasis, fibrosis, calcification and cardiomegaly). For external validation, the performance of DLAD-10 on a same-day computed tomography (CT)-confirmed dataset (normal:abnormal 53:147) and an open-source dataset (PadChest; normal:abnormal 339:334) was compared with that of three radiologists. Separate simulated reading tests were conducted on another dataset adjusted to real-world disease prevalence in the emergency department, consisting of four critical, 52 urgent and 146 nonurgent cases. Six radiologists participated in the simulated reading sessions with and without DLAD-10.DLAD-10 exhibited area under the receiver operating characteristic curve values of 0.895-1.00 in the CT-confirmed dataset and 0.913-0.997 in the PadChest dataset. DLAD-10 correctly classified significantly more critical abnormalities (95.0% (57/60)) than pooled radiologists (84.4% (152/180); p=0.01). In simulated reading tests for emergency department patients, pooled readers detected significantly more critical (70.8% (17/24) versus 29.2% (7/24); p=0.006) and urgent (82.7% (258/312) versus 78.2% (244/312); p=0.04) abnormalities when aided by DLAD-10. DLAD-10 assistance shortened the mean±sd time-to-report critical and urgent radiographs (640.5±466.3 versus 3371.0±1352.5 s and 1840.3±1141.1 versus 2127.1±1468.2 s, respectively; all p<0.01) and reduced the mean±sd interpretation time (20.5±22.8 versus 23.5±23.7 s; p<0.001).DLAD-10 showed excellent performance, improving radiologists' performance and shortening the reporting time for critical and urgent cases.

Deep learning reconstruction for contrast-enhanced CT of the upper abdomen: similar image quality with lower radiation dose in direct comparison with iterative reconstruction.

OBJECTIVE: To evaluate the effect of a commercial deep learning algorithm on the image quality of chest CT, focusing on the upper abdomen. METHODS: One hundred consecutive patients who simultaneously underwent contrast-enhanced chest and abdominal CT were collected. The radiation dose was optimized for each scan (mean CTDIvol: chest CT, 3.19 ± 1.53 mGy; abdominal CT, 7.10 ± 1.88 mGy). Three image sets were collected: chest CT reconstructed with an adaptive statistical iterative reconstruction (ASiR-CHT; 50% blending), chest CT with a deep learning algorithm (DLIR-CHT), and abdominal CT with ASiR (ASiR-ABD; 40% blending). Afterwards, the images covering the upper abdomen were extracted, and image noise, the signal-to-noise ratio (SNR), and the contrast-to-noise ratio (CNR) were measured. For subjective evaluation, three radiologists independently assessed noise, spatial resolution, presence of artifacts, and overall image quality. Additionally, readers selected the most preferable reconstruction technique among three image sets for each case. RESULTS: The average measured noise for DLIR-CHT, ASiR-CHT, and ASiR-ABD was 8.01 ± 2.81, 14.8 ± 2.56, and 12.3 ± 2.28, respectively (p < .001). Deep learning-based image reconstruction (DLIR) also showed the best SNR and CNR (p < .001). However, in the subjective analysis, ASiR-ABD showed less subjective noise than DLIR (2.94 ± 0.23 vs. 2.87 ± 0.26; p < .001), while DLIR showed better spatial resolution (2.60 ± 0.34 vs. 2.44 ± 0.31; p = .02). ASiR-ABD showed a better overall image quality (p = .001), but two of the three readers preferred DLIR more frequently. CONCLUSION: With < 50% of the radiation dose, DLIR chest CT showed comparable image quality in the upper abdomen to that of dedicated abdominal CT and was preferred by most readers. KEY POINTS: • With < 50% radiation dose, a deep learning algorithm applied to contrast-enhanced chest CT exhibited better image noise and signal-to-noise ratio than standard abdominal CT with the ASiR technique. • Pooled readers mostly preferred deep learning algorithm-reconstructed contrast-enhanced chest CT reconstructed using a standard ASiR-reconstructed abdominal CT. • Reconstruction algorithm-induced distortion artifacts were more frequently observed on deep learning algorithm-reconstructed images, but diagnostic difficulty was reported in only 0.3% of cases.

Image quality of ultralow-dose chest CT using deep learning techniques: potential superiority of vendor-agnostic post-processing over vendor-specific techniques.

OBJECTIVE: To compare the image quality between the vendor-agnostic and vendor-specific algorithms on ultralow-dose chest CT. METHODS: Vendor-agnostic deep learning post-processing model (DLM), vendor-specific deep learning image reconstruction (DLIR, high level), and adaptive statistical iterative reconstruction (ASiR, 70%) algorithms were employed. One hundred consecutive ultralow-dose noncontrast CT scans (CTDIvol; mean, 0.33 ± 0.056 mGy) were reconstructed with five algorithms: DLM-stnd (standard kernel), DLM-shrp (sharp kernel), DLIR, ASiR-stnd, and ASiR-shrp. Three thoracic radiologists blinded to the reconstruction algorithms reviewed five sets of 100 images and assessed subjective noise, spatial resolution, distortion artifact, and overall image quality. They selected the most preferred algorithm among five image sets for each case. Image noise and signal-to-noise ratio were measured. Edge-rise-distance was measured at a pulmonary vessel, i.e., the distance between two points where attenuation was 10% and 90% of maximal intravascular intensity. The skewness of attenuation was calculated in homogeneous areas. RESULTS: DLM-stnd, followed by DLIR, showed the best subjective noise on both lung and mediastinal windows, while DLIR yielded the least measured noise (ps < .0001). Compared to DLM-stnd, DLIR showed inferior subjective spatial resolution on lung window and higher edge-rise-distance (ps < .0001). Additionally, DLIR showed the most frequent distortion artifacts and deviated skewness (ps < .0001). DLM-stnd scored the best overall image quality, followed by DLM-shrp and DLIR (mean score 3.89 ± 0.19, 3.68 ± 0.24, and 3.53 ± 0.33; ps < .001). Two among three readers preferred DLM-stnd on both windows. CONCLUSION: Although DLIR provided the best quantitative noise profile, DLM-stnd showed the best overall image quality with fewer artifacts and was preferred by two among three readers. KEY POINTS: • A vendor-agnostic deep learning post-processing algorithm applied to ultralow-dose chest CT exhibited the best image quality compared to vendor-specific deep learning algorithm and ASiR techniques. • Two out of three readers preferred a vendor-agnostic deep learning post-processing algorithm in comparison to vendor-specific deep learning algorithm and ASiR techniques. • A vendor-specific deep learning reconstruction algorithm yielded the least image noise, but showed significantly more frequent specific distortion artifacts and increased skewness of attenuation compared to a vendor-agnostic algorithm.

Automatic pulmonary vessel segmentation on noncontrast chest CT: deep learning algorithm developed using spatiotemporally matched virtual noncontrast images and low-keV contrast-enhanced vessel maps.

OBJECTIVES: To develop a deep learning-based pulmonary vessel segmentation algorithm (DLVS) from noncontrast chest CT and to investigate its clinical implications in assessing vascular remodeling of chronic obstructive lung disease (COPD) patients. METHODS: For development, 104 pulmonary CT angiography scans (49,054 slices) using a dual-source CT were collected, and spatiotemporally matched virtual noncontrast and 50-keV images were generated. Vessel maps were extracted from the 50-keV images. The 3-dimensional U-Net-based DLVS was trained to segment pulmonary vessels (with a vessel map as the output) from virtual noncontrast images (as the input). For external validation, vendor-independent noncontrast CT images (n = 14) and the VESSEL 12 challenge open dataset (n = 3) were used. For each case, 200 points were selected including 20 intra-lesional points, and the probability value for each point was extracted. For clinical validation, we included 281 COPD patients with low-dose noncontrast CTs. The DLVS-calculated volume of vessels with a cross-sectional area < 5 mm2 (PVV5) and the PVV5 divided by total vessel volume (%PVV5) were measured. RESULTS: DLVS correctly segmented 99.1% of the intravascular points (1,387/1,400) and 93.1% of the extravascular points (1,309/1,400). The areas-under-the receiver-operating characteristic curve (AUROCs) were 0.977 and 0.969 for the two external validation datasets. For the COPD patients, both PPV5 and %PPV5 successfully differentiated severe patients whose FEV1 < 50 (AUROCs; 0.715 and 0.804) and were significantly correlated with the emphysema index (Ps < .05). CONCLUSIONS: DLVS successfully segmented pulmonary vessels on noncontrast chest CT by utilizing spatiotemporally matched 50-keV images from a dual-source CT scanner and showed promising clinical applicability in COPD. KEY POINTS: • We developed a deep learning pulmonary vessel segmentation algorithm using virtual noncontrast images and 50-keV enhanced images produced by a dual-source CT scanner. • Our algorithm successfully segmented vessels on diseased lungs. • Our algorithm showed promising results in assessing the loss of small vessel density in COPD patients.

Automatic prediction of left cardiac chamber enlargement from chest radiographs using convolutional neural network.

OBJECTIVE: To develop deep learning-based cardiac chamber enlargement-detection algorithms for left atrial (DLCE-LAE) and ventricular enlargement (DLCE-LVE), on chest radiographs METHODS: For training and internal validation of DLCE-LAE and -LVE, 5,045 chest radiographs (CRs; 2,463 normal and 2,393 LAE) and 1,012 CRs (456 normal and 456 LVE) matched with the same-day echocardiography were collected, respectively. External validation was performed using 107 temporally independent CRs. Reader performance test was conducted using the external validation dataset by five cardiothoracic radiologists without and with the results of DLCE. Classification performance of DLCE was evaluated and compared with those of the readers and conventional radiographic features, including cardiothoracic ratio, carinal angle, and double contour. In addition, DLCE-LAE was tested on 5,277 CRs from a healthcare screening program cohort. RESULTS: DLCE-LAE showed areas under the receiver operating characteristics curve (AUROCs) of 0.858 on external validation. On reader performance test, DLCE-LAE showed better results than pooled radiologists (AUROC 0.858 vs. 0.651; p < .001) and significantly increased their performance when used as a second reader (AUROC 0.651 vs. 0.722; p < .001). DLCE-LAE also showed a significantly higher AUROC than conventional radiographic findings (AUROC 0.858 vs. 0.535-0.706; all ps < .01). In the healthcare screening cohort, DLCE-LAE successfully detected 71.0% (142/200) CRs with moderate-to-severe LAE (93.5% [29/31] of severe cases), while yielding 11.8% (492/4,184) false-positive rate. DLCE-LVE showed AUROCs of 0.966 and 0.594 on internal and external validation, respectively. CONCLUSION: DLCE-LAE outperformed and improved cardiothoracic radiologists' performance in detecting LAE and showed promise in screening individuals with moderate-to-severe LAE in a healthcare screening cohort. KEY POINTS: • Our deep learning algorithm outperformed cardiothoracic radiologists in detecting left atrial enlargement on chest radiographs. • Cardiothoracic radiologists improved their performance in detecting left atrial enlargement when aided by the algorithm. • On a healthcare-screening cohort, our algorithm detected 71.0% (142/200) radiographs with moderate-to-severe left atrial enlargement while yielding 11.8% (492/4,184) false-positive rate.

Growth and Clinical Impact of 6-mm or Larger Subsolid Nodules after 5 Years of Stability at Chest CT.

Background It remains unclear whether 5 years of stability is sufficient to establish the benign behavior of subsolid nodules (SSNs) of the lung. There are no guidelines for the length of follow-up needed for these SSNs. Purpose To investigate the incidence of interval growth of pulmonary SSNs 6 mm or greater in diameter after 5 years of stability and their clinical outcome. Materials and Methods This retrospective study assessed SSNs 6 mm or greater that were stable for 5 years after detection (January 2002 to December 2018). The incidence of interval growth after 5 years of stability and the clinical and radiologic features of these SSNs were investigated. Clinical stage shifts of growing SSNs, presence of metastasis, and overall survival were assessed during the follow-up period. Subgroup analysis was performed in patients with nonenhanced thin-section (section thickness ≤1.5 mm) CT for interval growth after 5 years of stability. Results A total of 235 SSNs in 235 patients (mean age, 64 years ± 10 [standard deviation]; 132 women) were evaluated. There were 212 pure ground-glass nodules and 24 part-solid nodules. During follow-up (median, 112 months; range, 84-208 months), five of the 235 SSNs (2%; three primary ground-glass nodules and two part-solid nodules) showed interval growth. Three of these five growing SSNs were 10 mm or greater. Three of the five SSNs with interval growth had clinical stage shifts after growth (from Tis [in situ] to T1mi [minimally invasive] in one lesion; from T1mi to T1a in two lesions). There were no deaths or metastases from lung cancer during follow-up. Of 160 SSNs imaged with section thickness of 1.5 mm or less, two (1%) grew; both lesions were 10 mm or greater. Conclusion Only 2% of subsolid pulmonary nodules greater than or equal to 6 mm that had been stable for 5 years showed subsequent growth. At median follow-up of 9 years (after the initial 5-year period of stability), growth of those lung nodules had no clinical effect. © RSNA, 2020 See also the editorial by Naidich and Azour in this issue.

Deep learning algorithm for surveillance of pneumothorax after lung biopsy: a multicenter diagnostic cohort study.

OBJECTIVES: Pneumothorax is the most common and potentially life-threatening complication arising from percutaneous lung biopsy. We evaluated the performance of a deep learning algorithm for detection of post-biopsy pneumothorax in chest radiographs (CRs), in consecutive cohorts reflecting actual clinical situation. METHODS: We retrospectively included post-biopsy CRs of 1757 consecutive patients (1055 men, 702 women; mean age of 65.1 years) undergoing percutaneous lung biopsies from three institutions. A commercially available deep learning algorithm analyzed each CR to identify pneumothorax. We compared the performance of the algorithm with that of radiology reports made in the actual clinical practice. We also conducted a reader study, in which the performance of the algorithm was compared with those of four radiologists. Performances of the algorithm and radiologists were evaluated by area under receiver operating characteristic curves (AUROCs), sensitivity, and specificity, with reference standards defined by thoracic radiologists. RESULTS: Pneumothorax occurred in 17.5% (308/1757) of cases, out of which 16.6% (51/308) required catheter drainage. The AUROC, sensitivity, and specificity of the algorithm were 0.937, 70.5%, and 97.7%, respectively, for identification of pneumothorax. The algorithm exhibited higher sensitivity (70.2% vs. 55.5%, p < 0.001) and lower specificity (97.7% vs. 99.8%, p < 0.001), compared with those of radiology reports. In the reader study, the algorithm exhibited lower sensitivity (77.3% vs. 81.8-97.7%) and higher specificity (97.6% vs. 81.7-96.0%) than the radiologists. CONCLUSION: The deep learning algorithm appropriately identified pneumothorax in post-biopsy CRs in consecutive diagnostic cohorts. It may assist in accurate and timely diagnosis of post-biopsy pneumothorax in clinical practice. KEY POINTS: • A deep learning algorithm can identify chest radiographs with post-biopsy pneumothorax in multicenter consecutive cohorts reflecting actual clinical situation. • The deep learning algorithm has a potential role as a surveillance tool for accurate and timely diagnosis of post-biopsy pneumothorax.

Deep Learning for Chest Radiograph Diagnosis in the Emergency Department.

BackgroundThe performance of a deep learning (DL) algorithm should be validated in actual clinical situations, before its clinical implementation.PurposeTo evaluate the performance of a DL algorithm for identifying chest radiographs with clinically relevant abnormalities in the emergency department (ED) setting.Materials and MethodsThis single-center retrospective study included consecutive patients who visited the ED and underwent initial chest radiography between January 1 and March 31, 2017. Chest radiographs were analyzed with a commercially available DL algorithm. The performance of the algorithm was evaluated by determining the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity at predefined operating cutoffs (high-sensitivity and high-specificity cutoffs). The sensitivities and specificities of the algorithm were compared with those of the on-call radiology residents who interpreted the chest radiographs in the actual practice by using McNemar tests. If there were discordant findings between the algorithm and resident, the residents reinterpreted the chest radiographs by using the algorithm's output.ResultsA total of 1135 patients (mean age, 53 years ± 18; 582 men) were evaluated. In the identification of abnormal chest radiographs, the algorithm showed an AUC of 0.95 (95% confidence interval [CI]: 0.93, 0.96), a sensitivity of 88.7% (227 of 256 radiographs; 95% CI: 84.1%, 92.3%), and a specificity of 69.6% (612 of 879 radiographs; 95% CI: 66.5%, 72.7%) at the high-sensitivity cutoff and a sensitivity of 81.6% (209 of 256 radiographs; 95% CI: 76.3%, 86.2%) and specificity of 90.3% (794 of 879 radiographs; 95% CI: 88.2%, 92.2%) at the high-specificity cutoff. Radiology residents showed lower sensitivity (65.6% [168 of 256 radiographs; 95% CI: 59.5%, 71.4%], P < .001) and higher specificity (98.1% [862 of 879 radiographs; 95% CI: 96.9%, 98.9%], P < .001) compared with the algorithm. After reinterpretation of chest radiographs with use of the algorithm's outputs, the sensitivity of the residents improved (73.4% [188 of 256 radiographs; 95% CI: 68.0%, 78.8%], P = .003), whereas specificity was reduced (94.3% [829 of 879 radiographs; 95% CI: 92.8%, 95.8%], P < .001).ConclusionA deep learning algorithm used with emergency department chest radiographs showed diagnostic performance for identifying clinically relevant abnormalities and helped improve the sensitivity of radiology residents' evaluation.Published under a CC BY 4.0 license.Online supplemental material is available for this article.See also the editorial by Munera and Infante in this issue.

Development and Validation of Deep Learning-based Automatic Detection Algorithm for Malignant Pulmonary Nodules on Chest Radiographs.

Purpose To develop and validate a deep learning-based automatic detection algorithm (DLAD) for malignant pulmonary nodules on chest radiographs and to compare its performance with physicians including thoracic radiologists. Materials and Methods For this retrospective study, DLAD was developed by using 43 292 chest radiographs (normal radiograph-to-nodule radiograph ratio, 34 067:9225) in 34 676 patients (healthy-to-nodule ratio, 30 784:3892; 19 230 men [mean age, 52.8 years; age range, 18-99 years]; 15 446 women [mean age, 52.3 years; age range, 18-98 years]) obtained between 2010 and 2015, which were labeled and partially annotated by 13 board-certified radiologists, in a convolutional neural network. Radiograph classification and nodule detection performances of DLAD were validated by using one internal and four external data sets from three South Korean hospitals and one U.S. hospital. For internal and external validation, radiograph classification and nodule detection performances of DLAD were evaluated by using the area under the receiver operating characteristic curve (AUROC) and jackknife alternative free-response receiver-operating characteristic (JAFROC) figure of merit (FOM), respectively. An observer performance test involving 18 physicians, including nine board-certified radiologists, was conducted by using one of the four external validation data sets. Performances of DLAD, physicians, and physicians assisted with DLAD were evaluated and compared. Results According to one internal and four external validation data sets, radiograph classification and nodule detection performances of DLAD were a range of 0.92-0.99 (AUROC) and 0.831-0.924 (JAFROC FOM), respectively. DLAD showed a higher AUROC and JAFROC FOM at the observer performance test than 17 of 18 and 15 of 18 physicians, respectively (P < .05), and all physicians showed improved nodule detection performances with DLAD (mean JAFROC FOM improvement, 0.043; range, 0.006-0.190; P < .05). Conclusion This deep learning-based automatic detection algorithm outperformed physicians in radiograph classification and nodule detection performance for malignant pulmonary nodules on chest radiographs, and it enhanced physicians' performances when used as a second reader. © RSNA, 2018 Online supplemental material is available for this article.

Value of virtual monochromatic spectral image of dual-layer spectral detector CT with noise reduction algorithm for image quality improvement in obese simulated body phantom.

BACKGROUND: Dual-layer spectral detector CT (SDCT) may provide several theoretical advantages over pre-existing DECT approaches in terms of adjustment-free sampling number and dose modulation, beam hardening correction, and production spectral images by post-processing. In addition, by adopting noise reduction algorithm, high contrast resolution was expected even in low keV level. We surmised that this improvement would be beneficial to obese people. Therefore, our aim of study is to compare image quality of virtual monochromatic spectral images (VMI) and polychromatic images reconstructed from SDCT with different body size and radiation dose using anthropomorphic liver phantom. METHODS: One small and one large size of body phantoms, each containing eight (four high- and four low-contrast) simulated focal liver lesions (FLLs) were scanned by SDCT (at 120 kVp) using different Dose Right Indexes (DRIs). VMI were reconstructed from spectral base images from 40 keV to 200 keV. Hybrid iterative reconstruction (iDose4) was used for polychromatic image reconstruction. Image noise and contrast to noise ratio (CNR) were compared. Five radiologists independently rated lesion conspicuity, diagnostic acceptability and subjective noise level in every image sets, and determined optimal keV level in VMI. RESULTS: Compare with conventional polychromatic images, VMI showed superior CNR at low keV level regardless of phantom size at every examined DRIs (Ps < 0.05). As body size increased, VMI had more gradual CNR decrease and noise increase than conventional polychromatic images. For low contrast FLLs in large phantom, lesion conspicuities at low radiation dose levels (DRI 16 and 19) were significantly increased in VMI (Ps < 0.05). Subjective image noise and diagnostic acceptabilities were significantly improved at VMI in both phantom size. CONCLUSIONS: VMI of dual-layer spectral detector CT with noise reduction algorithm provides improved CNR, noise reduction, and better subjective image quality in imaging of obese simulated liver phantom compared with polychromatic images. This may hold promise for improving detection of liver lesions and improved imaging of obese patients.