Full Peer Review Research Articles

WAW-TACE: A Hepatocellular Carcinoma Multiphase CT Dataset with Segmentations, Radiomics Features, and Clinical Data.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. The WAW-TACE dataset contains baseline multiphase abdominal CT images from 233 treatment-naive patients with hepatocellular carcinoma treated with transarterial chemoembolization and includes with 377 hand-crafted liver tumor masks, automated segmentations of multiple internal organs, extracted radiomics features, and corresponding extensive clinical data. The dataset can be accessed at: https://zenodo.org/records/12741586 (DOI:10.5281/zenodo.11063784).

The RSNA Abdominal Traumatic Injury CT (RATIC) Dataset.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. The RSNA Abdominal Traumatic Injury CT (RATIC) dataset contains 4,274 abdominal CT studies with annotations related to traumatic injuries and is available at https://imaging.rsna.org/dataset/5 and https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection. ©RSNA, 2024.

Addressing the Generalizability of AI in Radiology Using a Novel Data Augmentation Framework with Synthetic Patient Image Data: Proof-of-Concept and External Validation Classification Tasks in Multiple Sclerosis.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Artificial intelligence (AI) models often face performance drops after deployment to external datasets. This study evaluated the potential of a novel data augmentation framework based on generative adversarial networks (GAN) that creates synthetic patient image data during model training to improve model generalizability. Model development and external testing were performed for a given classification task, namely the detection of new fluid-attenuated inversion recovery (FLAIR) lesions on MRI during longitudinal follow-up of patients with multiple sclerosis (MS). An internal dataset of 669 patients with MS (n = 3083 examinations) was used to develop an attention-based network, trained both with and without the inclusion of the GAN-based synthetic data augmentation framework. External testing was performed on 134 patients with MS from a different institution, with MR images acquired using different scanners and protocols than images used during training. Models trained using synthetic data augmentation showed a significant performance improvement when applied on external data (AUC 83.6% without synthetic data versus AUC 93.3% with synthetic data augmentation, P = .03), achieving comparable results to the internal test set (AUC 95.5%, P = .53), whereas models without synthetic data augmentation demonstrated a performance drop upon external testing (AUC 93.8% on internal dataset versus AUC 83.6% on external data, P = .03). Data augmentation with synthetic patient data substantially improved performance of AI models on unseen MRI data and may be extended to other clinical conditions or tasks to mitigate domain shift, limit class imbalance, and enhance the robustness of AI applications in medical imaging. ©RSNA, 2024.

Boosting Deep Learning for Interpretable Brain MRI Lesion Detection through the Integration of Radiology Report Information.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To guide a deep learning (DL) model's attention toward brain lesion MRI characteristics by incorporating radiology report-derived textual features to achieve interpretable lesion detection. Materials and Methods In this retrospective study, 35282 brain MRI scans (January 2018-June 2023) and corresponding radiology reports from center 1 were used for training, validation, and internal testing. 2655 brain MRI scans (January 2022-December 2022) from centers 2-5 were reserved for external testing. Textual features were extracted from radiology reports to guide a DL model (ReportGuidedNet) focusing on lesion characteristics. Another DL model (PlainNet) without textual features was developed for comparative analysis. Both models diagnosed 15 conditions, including 14 diseases and normal brains. Performance of each model was assessed by calculating macro-and microaveraged area under the receiver operating characteristic curves (ma-AUC, mi-AUC). Attention maps, visualizing model attention, were assessed with a 5-point Likert scale. Results ReportGuidedNet outperformed PlainNet for all diagnoses on both internal (ma-AUC: 0.93 [95% CI: 0.91- 0.95] versus 0.85 [95% CI: 0.81-0.88]; mi-AUC: 0.93 [95% CI: 0.90-0.95] versus 0.89 [95% CI: 0.83-0.92]) and external (ma-AUC: 0.91 [95% CI: 0.88-0.93] versus 0.75 [95% CI: 0.72-0.79]; mi-AUC: 0.90 [95% CI: 0.87-0.92] versus 0.76 [95% CI: 0.72-0.80]) testing sets. The performance difference between internal and external testing sets was smaller for ReportGuidedNet than for PlainNet (Δma-AUC: 0.03 versus 0.10; Δmi-AUC: 0.02 versus 0.13). The Likert scale score of ReportGuidedNet was higher than that of PlainNet (mean ± SD: 2.50 ± 1.09 versus 1.32 ± 1.20; P < .001). Conclusion The integration of radiology report textual features improved the DL model's ability to detect brain lesions, enhancing interpretability and generalizability. Published under a CC BY 4.0 license.

AI-integrated Screening to Replace Double Reading of Mammograms: A Population-wide Accuracy and Feasibility Study.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Mammography screening supported by deep learning-based artificial intelligence (AI) solutions can potentially reduce workload without compromising breast cancer detection accuracy, but the site of deployment in the workflow might be crucial. This retrospective study compared three simulated AI-integrated screening scenarios with standard double reading with arbitration in a sample of 249,402 mammograms from a representative screening population. A commercial AI system replaced the first reader (Scenario 1: Integrated AIfirst), the second reader (Scenario 2: Integrated AIsecond), or both readers for triaging of low- and high-risk cases (Integrated AItriage). AI threshold values were partly chosen based on previous validation and fixing screen-read volume reduction at approximately 50% across scenarios. Detection accuracy measures were calculated. Compared with standard double reading, Integrated AIfirst showed no evidence of a difference in accuracy metrics except for a higher arbitration rate (+0.99%; P < .001). Integrated AIsecond had lower sensitivity (-1.58%; P < 0.001), negative predictive value (NPV) (- 0.01%; P < .001) and recall rate (< 0.06%; P = 0.04), but a higher positive predictive value (PPV) (+0.03%; P < .001) and arbitration rate (+1.22%; P < .001). Integrated AItriage achieved higher sensitivity (+1.33%; P < .001), PPV (+0.36%; P = .03), and NPV (+0.01%; P < .001) but lower arbitration rate (-0.88%; P < .001). Replacing one or both readers with AI seems feasible, however, the site of application in the workflow can have clinically relevant effects on accuracy and workload. ©RSNA, 2024.

Deep Learning-based Unsupervised Domain Adaptation via a Unified Model for Prostate Lesion Detection Using Multisite Biparametric MRI Datasets.

Purpose To determine whether the unsupervised domain adaptation (UDA) method with generated images improves the performance of a supervised learning (SL) model for prostate cancer (PCa) detection using multisite biparametric (bp) MRI datasets. Materials and Methods This retrospective study included data from 5150 patients (14 191 samples) collected across nine different imaging centers. A novel UDA method using a unified generative model was developed for PCa detection using multisite bpMRI datasets. This method translates diffusion-weighted imaging (DWI) acquisitions, including apparent diffusion coefficient (ADC) and individual diffusion-weighted (DW) images acquired using various b values, to align with the style of images acquired using b values recommended by Prostate Imaging Reporting and Data System (PI-RADS) guidelines. The generated ADC and DW images replace the original images for PCa detection. An independent set of 1692 test cases (2393 samples) was used for evaluation. The area under the receiver operating characteristic curve (AUC) was used as the primary metric, and statistical analysis was performed via bootstrapping. Results For all test cases, the AUC values for baseline SL and UDA methods were 0.73 and 0.79 (P < .001), respectively, for PCa lesions with PI-RADS score of 3 or greater and 0.77 and 0.80 (P < .001) for lesions with PI-RADS scores of 4 or greater. In the 361 test cases under the most unfavorable image acquisition setting, the AUC values for baseline SL and UDA were 0.49 and 0.76 (P < .001) for lesions with PI-RADS scores of 3 or greater and 0.50 and 0.77 (P < .001) for lesions with PI-RADS scores of 4 or greater. Conclusion UDA with generated images improved the performance of SL methods in PCa lesion detection across multisite datasets with various b values, especially for images acquired with significant deviations from the PI-RADS-recommended DWI protocol (eg, with an extremely high b value). Keywords: Prostate Cancer Detection, Multisite, Unsupervised Domain Adaptation, Diffusion-weighted Imaging, b Value Supplemental material is available for this article. © RSNA, 2024.

Presurgical Upgrade Prediction of DCIS to Invasive Ductal Carcinoma Using Time-dependent Deep Learning Models with DCE MRI.

Purpose To determine whether time-dependent deep learning models can outperform single time point models in predicting preoperative upgrade of ductal carcinoma in situ (DCIS) to invasive malignancy at dynamic contrast-enhanced (DCE) breast MRI without a lesion segmentation prerequisite. Materials and Methods In this exploratory study, 154 cases of biopsy-proven DCIS (25 upgraded at surgery and 129 not upgraded) were selected consecutively from a retrospective cohort of preoperative DCE MRI in women with a mean age of 59 years at time of diagnosis from 2012 to 2022. Binary classification was implemented with convolutional neural network (CNN)-long short-term memory (LSTM) architectures benchmarked against traditional CNNs without manual segmentation of the lesions. Combinatorial performance analysis of ResNet50 versus VGG16-based models was performed with each contrast phase. Binary classification area under the receiver operating characteristic curve (AUC) was reported. Results VGG16-based models consistently provided better holdout test AUCs than did ResNet50 in CNN and CNN-LSTM studies (multiphase test AUC, 0.67 vs 0.59, respectively, for CNN models [P = .04] and 0.73 vs 0.62 for CNN-LSTM models [P = .008]). The time-dependent model (CNN-LSTM) provided a better multiphase test AUC over single time point (CNN) models (0.73 vs 0.67; P = .04). Conclusion Compared with single time point architectures, sequential deep learning algorithms using preoperative DCE MRI improved prediction of DCIS lesions upgraded to invasive malignancy without the need for lesion segmentation. Keywords: MRI, Dynamic Contrast-enhanced, Breast, Convolutional Neural Network (CNN) Supplemental material is available for this article. © RSNA, 2024.

Deep Learning Segmentation of Ascites on Abdominal CT Scans for Automatic Volume Quantification.

Purpose To evaluate the performance of an automated deep learning method in detecting ascites and subsequently quantifying its volume in patients with liver cirrhosis and patients with ovarian cancer. Materials and Methods This retrospective study included contrast-enhanced and noncontrast abdominal-pelvic CT scans of patients with cirrhotic ascites and patients with ovarian cancer from two institutions, National Institutes of Health (NIH) and University of Wisconsin (UofW). The model, trained on The Cancer Genome Atlas Ovarian Cancer dataset (mean age [±SD], 60 years ± 11; 143 female), was tested on two internal datasets (NIH-LC and NIH-OV) and one external dataset (UofW-LC). Its performance was measured by the F1/Dice coefficient, SDs, and 95% CIs, focusing on ascites volume in the peritoneal cavity. Results On NIH-LC (25 patients; mean age, 59 years ± 14; 14 male) and NIH-OV (166 patients; mean age, 65 years ± 9; all female), the model achieved F1/Dice scores of 85.5% ± 6.1 (95% CI: 83.1, 87.8) and 82.6% ± 15.3 (95% CI: 76.4, 88.7), with median volume estimation errors of 19.6% (IQR, 13.2%-29.0%) and 5.3% (IQR: 2.4%-9.7%), respectively. On UofW-LC (124 patients; mean age, 46 years ± 12; 73 female), the model had a F1/Dice score of 83.0% ± 10.7 (95% CI: 79.8, 86.3) and median volume estimation error of 9.7% (IQR, 4.5%-15.1%). The model showed strong agreement with expert assessments, with r2 values of 0.79, 0.98, and 0.97 across the test sets. Conclusion The proposed deep learning method performed well in segmenting and quantifying the volume of ascites in patients with cirrhosis and those with ovarian cancer, in concordance with expert radiologist assessments. Keywords: Abdomen/GI, Cirrhosis, Deep Learning, Segmentation Supplemental material is available for this article. © RSNA, 2024 See also commentary by Aisen and Rodrigues in this issue.

Improving Fairness of Automated Chest Radiograph Diagnosis by Contrastive Learning.

Purpose To develop an artificial intelligence model that uses supervised contrastive learning (SCL) to minimize bias in chest radiograph diagnosis. Materials and Methods In this retrospective study, the proposed method was evaluated on two datasets: the Medical Imaging and Data Resource Center (MIDRC) dataset with 77 887 chest radiographs in 27 796 patients collected as of April 20, 2023, for COVID-19 diagnosis and the National Institutes of Health ChestX-ray14 dataset with 112 120 chest radiographs in 30 805 patients collected between 1992 and 2015. In the ChestX-ray14 dataset, thoracic abnormalities included atelectasis, cardiomegaly, effusion, infiltration, mass, nodule, pneumonia, pneumothorax, consolidation, edema, emphysema, fibrosis, pleural thickening, and hernia. The proposed method used SCL with carefully selected positive and negative samples to generate fair image embeddings, which were fine-tuned for subsequent tasks to reduce bias in chest radiograph diagnosis. The method was evaluated using the marginal area under the receiver operating characteristic curve difference (∆mAUC). Results The proposed model showed a significant decrease in bias across all subgroups compared with the baseline models, as evidenced by a paired t test (P < .001). The ∆mAUCs obtained by the proposed method were 0.01 (95% CI: 0.01, 0.01), 0.21 (95% CI: 0.21, 0.21), and 0.10 (95% CI: 0.10, 0.10) for sex, race, and age subgroups, respectively, on the MIDRC dataset and 0.01 (95% CI: 0.01, 0.01) and 0.05 (95% CI: 0.05, 0.05) for sex and age subgroups, respectively, on the ChestX-ray14 dataset. Conclusion Employing SCL can mitigate bias in chest radiograph diagnosis, addressing concerns of fairness and reliability in deep learning-based diagnostic methods. Keywords: Thorax, Diagnosis, Supervised Learning, Convolutional Neural Network (CNN), Computer-aided Diagnosis (CAD) Supplemental material is available for this article. © RSNA, 2024 See also the commentary by Johnson in this issue.

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

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To develop a highly generalizable weakly supervised model to automatically detect and localize image- level intracranial hemorrhage (ICH) using study-level labels. Materials and Methods In this retrospective study, the proposed model was pretrained on the image-level RSNA dataset and fine-tuned on a local dataset using attention-based bidirectional long-short-term memory networks. This local training dataset included 10,699 noncontrast head CT scans from 7469 patients with ICH study-level labels extracted from radiology reports. Model performance was compared with that of two senior neuroradiologists on 100 random test scans using the McNemar test, and its generalizability was evaluated on an external independent dataset. Results The model achieved a positive predictive value (PPV) of 85.7% (95% CI: [84.0%, 87.4%]) and an AUC of 0.96 (95% CI: [0.96, 0.97]) on the held-out local test set (n = 7243, 3721 female) and 89.3% (95% CI: [87.8%, 90.7%]) and 0.96 (95% CI: [0.96, 0.97]), respectively, on the external test set (n = 491, 178 female). For 100 randomly selected samples, the model achieved performance on par with two neuroradiologists, but with a significantly faster (P < .05) diagnostic time of 5.04 seconds per scan (versus 86 seconds and 22.2 seconds for the two neuroradiologists, respectively). The model's attention weights and heatmaps visually aligned with neuroradiologists' interpretations. Conclusion The proposed model demonstrated high generalizability and high PPVs, offering a valuable tool for expedited ICH detection and prioritization while reducing false-positive interruptions in radiologists' workflows. ©RSNA, 2024.

Improving Computer-aided Detection for Digital Breast Tomosynthesis by Incorporating Temporal Change.

Purpose To develop a deep learning algorithm that uses temporal information to improve the performance of a previously published framework of cancer lesion detection for digital breast tomosynthesis. Materials and Methods This retrospective study analyzed the current and the 1-year-prior Hologic digital breast tomosynthesis screening examinations from eight different institutions between 2016 and 2020. The dataset contained 973 cancer and 7123 noncancer cases. The front end of this algorithm was an existing deep learning framework that performed single-view lesion detection followed by ipsilateral view matching. For this study, PriorNet was implemented as a cascaded deep learning module that used the additional growth information to refine the final probability of malignancy. Data from seven of the eight sites were used for training and validation, while the eighth site was reserved for external testing. Model performance was evaluated using localization receiver operating characteristic curves. Results On the validation set, PriorNet showed an area under the receiver operating characteristic curve (AUC) of 0.931 (95% CI: 0.930, 0.931), which outperformed both baseline models using single-view detection (AUC, 0.892 [95% CI: 0.891, 0.892]; P < .001) and ipsilateral matching (AUC, 0.915 [95% CI: 0.914, 0.915]; P < .001). On the external test set, PriorNet achieved an AUC of 0.896 (95% CI: 0.885, 0.896), outperforming both baselines (AUC, 0.846 [95% CI: 0.846, 0.847]; P < .001 and AUC, 0.865 [95% CI: 0.865, 0.866]; P < .001, respectively). In the high sensitivity range of 0.9 to 1.0, the partial AUC of PriorNet was significantly higher (P < .001) relative to both baselines. Conclusion PriorNet using temporal information further improved the breast cancer detection performance of an existing digital breast tomosynthesis cancer detection framework. Keywords: Digital Breast Tomosynthesis, Computer-aided Detection, Breast Cancer, Deep Learning © RSNA, 2024 See also commentary by Lee in this issue.

Anatomy-specific Progression Classification in Chest Radiographs via Weakly Supervised Learning.

Purpose To develop a machine learning approach for classifying disease progression in chest radiographs using weak labels automatically derived from radiology reports. Materials and Methods In this retrospective study, a twin neural network was developed to classify anatomy-specific disease progression into four categories: improved, unchanged, worsened, and new. A two-step weakly supervised learning approach was employed, pretraining the model on 243 008 frontal chest radiographs from 63 877 patients (mean age, 51.7 years ± 17.0 [SD]; 34 813 [55%] female) included in the MIMIC-CXR database and fine-tuning it on the subset with progression labels derived from consecutive studies. Model performance was evaluated for six pathologic observations on test datasets of unseen patients from the MIMIC-CXR database. Area under the receiver operating characteristic (AUC) analysis was used to evaluate classification performance. The algorithm is also capable of generating bounding-box predictions to localize areas of new progression. Recall, precision, and mean average precision were used to evaluate the new progression localization. One-tailed paired t tests were used to assess statistical significance. Results The model outperformed most baselines in progression classification, achieving macro AUC scores of 0.72 ± 0.004 for atelectasis, 0.75 ± 0.007 for consolidation, 0.76 ± 0.017 for edema, 0.81 ± 0.006 for effusion, 0.7 ± 0.032 for pneumonia, and 0.69 ± 0.01 for pneumothorax. For new observation localization, the model achieved mean average precision scores of 0.25 ± 0.03 for atelectasis, 0.34 ± 0.03 for consolidation, 0.33 ± 0.03 for edema, and 0.31 ± 0.03 for pneumothorax. Conclusion Disease progression classification models were developed on a large chest radiograph dataset, which can be used to monitor interval changes and detect new pathologic conditions on chest radiographs. Keywords: Prognosis, Unsupervised Learning, Transfer Learning, Convolutional Neural Network (CNN), Emergency Radiology, Named Entity Recognition Supplemental material is available for this article. © RSNA, 2024 See also commentary by Alves and Venkadesh in this issue.

Deep Learning to Detect Intracranial Hemorrhage in a National Teleradiology Program and the Impact on Interpretation Time.

The diagnostic performance of an artificial intelligence (AI) clinical decision support solution for acute intracranial hemorrhage (ICH) detection was assessed in a large teleradiology practice. The impact on radiologist read times and system efficiency was also quantified. A total of 61 704 consecutive noncontrast head CT examinations were retrospectively evaluated. System performance was calculated along with mean and median read times for CT studies obtained before (baseline, pre-AI period; August 2021 to May 2022) and after (post-AI period; January 2023 to February 2024) AI implementation. The AI solution had a sensitivity of 75.6%, specificity of 92.1%, accuracy of 91.7%, prevalence of 2.70%, and positive predictive value of 21.1%. Of the 56 745 post-AI CT scans with no bleed identified by a radiologist, examinations falsely flagged as suspected ICH by the AI solution (n = 4464) took an average of 9 minutes 40 seconds (median, 8 minutes 7 seconds) to interpret as compared with 8 minutes 25 seconds (median, 6 minutes 48 seconds) for unremarkable CT scans before AI (n = 49 007) (P < .001) and 8 minutes 38 seconds (median, 6 minutes 53 seconds) after AI when ICH was not suspected by the AI solution (n = 52 281) (P < .001). CT scans with no bleed identified by the AI but reported as positive for ICH by the radiologist (n = 384) took an average of 14 minutes 23 seconds (median, 13 minutes 35 seconds) to interpret as compared with 13 minutes 34 seconds (median, 12 minutes 30 seconds) for CT scans correctly reported as a bleed by the AI (n = 1192) (P = .04). With lengthened read times for falsely flagged examinations, system inefficiencies may outweigh the potential benefits of using the tool in a high volume, low prevalence environment. Keywords: Artificial Intelligence, Intracranial Hemorrhage, Read Time, Report Turnaround Time, System Efficiency Supplemental material is available for this article. © RSNA, 2024.

Open Access Data and Deep Learning for Cardiac Device Identification on Standard DICOM and Smartphone-based Chest Radiographs.

Purpose To develop and evaluate a publicly available deep learning model for segmenting and classifying cardiac implantable electronic devices (CIEDs) on Digital Imaging and Communications in Medicine (DICOM) and smartphone-based chest radiographs. Materials and Methods This institutional review board-approved retrospective study included patients with implantable pacemakers, cardioverter defibrillators, cardiac resynchronization therapy devices, and cardiac monitors who underwent chest radiography between January 2012 and January 2022. A U-Net model with a ResNet-50 backbone was created to classify CIEDs on DICOM and smartphone images. Using 2321 chest radiographs in 897 patients (median age, 76 years [range, 18-96 years]; 625 male, 272 female), CIEDs were categorized into four manufacturers, 27 models, and one "other" category. Five smartphones were used to acquire 11 072 images. Performance was reported using the Dice coefficient on the validation set for segmentation or balanced accuracy on the test set for manufacturer and model classification, respectively. Results The segmentation tool achieved a mean Dice coefficient of 0.936 (IQR: 0.890-0.958). The model had an accuracy of 94.36% (95% CI: 90.93%, 96.84%; 251 of 266) for CIED manufacturer classification and 84.21% (95% CI: 79.31%, 88.30%; 224 of 266) for CIED model classification. Conclusion The proposed deep learning model, trained on both traditional DICOM and smartphone images, showed high accuracy for segmentation and classification of CIEDs on chest radiographs. Keywords: Conventional Radiography, Segmentation Supplemental material is available for this article. © RSNA, 2024 See also the commentary by Júdice de Mattos Farina and Celi in this issue.

Artificial Intelligence Outcome Prediction in Neonates with Encephalopathy (AI-OPiNE).

Purpose To develop a deep learning algorithm to predict 2-year neurodevelopmental outcomes in neonates with hypoxic-ischemic encephalopathy using MRI and basic clinical data. Materials and Methods In this study, MRI data of term neonates with encephalopathy in the High-dose Erythropoietin for Asphyxia and Encephalopathy (HEAL) trial (ClinicalTrials.gov: NCT02811263), who were enrolled from 17 institutions between January 25, 2017, and October 9, 2019, were retrospectively analyzed. The harmonized MRI protocol included T1-weighted, T2-weighted, and diffusion tensor imaging. Deep learning classifiers were trained to predict the primary outcome of the HEAL trial (death or any neurodevelopmental impairment at 2 years) using multisequence MRI and basic clinical variables, including sex and gestational age at birth. Model performance was evaluated on test sets comprising 10% of cases from 15 institutions (in-distribution test set, n = 41) and 10% of cases from two institutions (out-of-distribution test set, n = 41). Model performance in predicting additional secondary outcomes, including death alone, was also assessed. Results For the 414 neonates (mean gestational age, 39 weeks ± 1.4 [SD]; 232 male, 182 female), in the study cohort, 198 (48%) died or had any neurodevelopmental impairment at 2 years. The deep learning model achieved an area under the receiver operating characteristic curve (AUC) of 0.74 (95% CI: 0.60, 0.86) and 63% accuracy in the in-distribution test set and an AUC of 0.77 (95% CI: 0.63, 0.90) and 78% accuracy in the out-of-distribution test set. Performance was similar or better for predicting secondary outcomes. Conclusion Deep learning analysis of neonatal brain MRI yielded high performance for predicting 2-year neurodevelopmental outcomes. Keywords: Convolutional Neural Network (CNN), Prognosis, Pediatrics, Brain, Brain Stem Clinical trial registration no. NCT02811263 Supplemental material is available for this article. © RSNA, 2024 See also commentary by Rafful and Reis Teixeira in this issue.

Clinical, Cultural, Computational, and Regulatory Considerations to Deploy AI in Radiology: Perspectives of RSNA and MICCAI Experts.

The Radiological Society of North of America (RSNA) and the Medical Image Computing and Computer Assisted Intervention (MICCAI) Society have led a series of joint panels and seminars focused on the present impact and future directions of artificial intelligence (AI) in radiology. These conversations have collected viewpoints from multidisciplinary experts in radiology, medical imaging, and machine learning on the current clinical penetration of AI technology in radiology and how it is impacted by trust, reproducibility, explainability, and accountability. The collective points-both practical and philosophical-define the cultural changes for radiologists and AI scientists working together and describe the challenges ahead for AI technologies to meet broad approval. This article presents the perspectives of experts from MICCAI and RSNA on the clinical, cultural, computational, and regulatory considerations-coupled with recommended reading materials-essential to adopt AI technology successfully in radiology and, more generally, in clinical practice. The report emphasizes the importance of collaboration to improve clinical deployment, highlights the need to integrate clinical and medical imaging data, and introduces strategies to ensure smooth and incentivized integration. Keywords: Adults and Pediatrics, Computer Applications-General (Informatics), Diagnosis, Prognosis © RSNA, 2024.

Deep Learning for Breast Cancer Risk Prediction: Application to a Large Representative UK Screening Cohort.

Purpose To develop an artificial intelligence (AI) deep learning tool capable of predicting future breast cancer risk from a current negative screening mammographic examination and to evaluate the model on data from the UK National Health Service Breast Screening Program. Materials and Methods The OPTIMAM Mammography Imaging Database contains screening data, including mammograms and information on interval cancers, for more than 300 000 female patients who attended screening at three different sites in the United Kingdom from 2012 onward. Cancer-free screening examinations from women aged 50-70 years were performed and classified as risk-positive or risk-negative based on the occurrence of cancer within 3 years of the original examination. Examinations with confirmed cancer and images containing implants were excluded. From the resulting 5264 risk-positive and 191 488 risk-negative examinations, training (n = 89 285), validation (n = 2106), and test (n = 39 351) datasets were produced for model development and evaluation. The AI model was trained to predict future cancer occurrence based on screening mammograms and patient age. Performance was evaluated on the test dataset using the area under the receiver operating characteristic curve (AUC) and compared across subpopulations to assess potential biases. Interpretability of the model was explored, including with saliency maps. Results On the hold-out test set, the AI model achieved an overall AUC of 0.70 (95% CI: 0.69, 0.72). There was no evidence of a difference in performance across the three sites, between patient ethnicities, or across age groups. Visualization of saliency maps and sample images provided insights into the mammographic features associated with AI-predicted cancer risk. Conclusion The developed AI tool showed good performance on a multisite, United Kingdom-specific dataset. Keywords: Deep Learning, Artificial Intelligence, Breast Cancer, Screening, Risk Prediction Supplemental material is available for this article. ©RSNA, 2024.

Optimizing Performance of Transformer-based Models for Fetal Brain MR Image Segmentation.

Purpose To test the performance of a transformer-based model when manipulating pretraining weights, dataset size, and input size and comparing the best model with the reference standard and state-of-the-art models for a resting-state functional (rs-fMRI) fetal brain extraction task. Materials and Methods An internal retrospective dataset (172 fetuses, 519 images; collected 2018-2022) was used to investigate influence of dataset size, pretraining approaches, and image input size on Swin-U-Net transformer (UNETR) and UNETR models. The internal and external (131 fetuses, 561 images) datasets were used to cross-validate and to assess generalization capability of the best model versus state-of-the-art models on different scanner types and number of gestational weeks (GWs). The Dice similarity coefficient (DSC) and the balanced average Hausdorff distance (BAHD) were used as segmentation performance metrics. Generalized equation estimation multifactorial models were used to assess significant model and interaction effects of interest. Results The Swin-UNETR model was not affected by the pretraining approach and dataset size and performed best with the mean dataset image size, with a mean DSC of 0.92 and BAHD of 0.097. Swin-UNETR was not affected by scanner type. Generalization results on the internal dataset showed that Swin-UNETR had lower performance compared with the reference standard models and comparable performance on the external dataset. Cross-validation on internal and external test sets demonstrated better and comparable performance of Swin-UNETR versus convolutional neural network architectures during the late-fetal period (GWs > 25) but lower performance during the midfetal period (GWs ≤ 25). Conclusion Swin-UNTER showed flexibility in dealing with smaller datasets, regardless of pretraining approaches. For fetal brain extraction from rs-fMR images, Swin-UNTER showed comparable performance with that of reference standard models during the late-fetal period and lower performance during the early GW period. Keywords: Transformers, CNN, Medical Imaging Segmentation, MRI, Dataset Size, Input Size, Transfer Learning Supplemental material is available for this article. © RSNA, 2024.

Radiomics Beyond the Hype: A Critical Evaluation Toward Oncologic Clinical Use.

Radiomics is a promising and fast-developing field within oncology that involves the mining of quantitative high-dimensional data from medical images. Radiomics has the potential to transform cancer management, whereby radiomics data can be used to aid early tumor characterization, prognosis, risk stratification, treatment planning, treatment response assessment, and surveillance. Nevertheless, certain challenges have delayed the clinical adoption and acceptability of radiomics in routine clinical practice. The objectives of this report are to (a) provide a perspective on the translational potential and potential impact of radiomics in oncology; (b) explore frequent challenges and mistakes in its derivation, encompassing study design, technical requirements, standardization, model reproducibility, transparency, data sharing, privacy concerns, quality control, as well as the complexity of multistep processes resulting in less radiologist-friendly interfaces; (c) discuss strategies to overcome these challenges and mistakes; and (d) propose measures to increase the clinical use and acceptability of radiomics, taking into account the different perspectives of patients, health care workers, and health care systems. Keywords: Radiomics, Oncology, Cancer Management, Artificial Intelligence © RSNA, 2024.

Impact of Transfer Learning Using Local Data on Performance of a Deep Learning Model for Screening Mammography.

Purpose To investigate the issues of generalizability and replication of deep learning models by assessing performance of a screening mammography deep learning system developed at New York University (NYU) on a local Australian dataset. Materials and Methods In this retrospective study, all individuals with biopsy or surgical pathology-proven lesions and age-matched controls were identified from a South Australian public mammography screening program (January 2010 to December 2016). The primary outcome was deep learning system performance-measured with area under the receiver operating characteristic curve (AUC)-in classifying invasive breast cancer or ductal carcinoma in situ (n = 425) versus no malignancy (n = 490) or benign lesions (n = 44). The NYU system, including models without (NYU1) and with (NYU2) heatmaps, was tested in its original form, after training from scratch (without transfer learning), and after retraining with transfer learning. Results The local test set comprised 959 individuals (mean age, 62.5 years ± 8.5 [SD]; all female). The original AUCs for the NYU1 and NYU2 models were 0.83 (95% CI: 0.82, 0.84) and 0.89 (95% CI: 0.88, 0.89), respectively. When NYU1 and NYU2 were applied in their original form to the local test set, the AUCs were 0.76 (95% CI: 0.73, 0.79) and 0.84 (95% CI: 0.82, 0.87), respectively. After local training without transfer learning, the AUCs were 0.66 (95% CI: 0.62, 0.69) and 0.86 (95% CI: 0.84, 0.88). After retraining with transfer learning, the AUCs were 0.82 (95% CI: 0.80, 0.85) and 0.86 (95% CI: 0.84, 0.88). Conclusion A deep learning system developed using a U.S. dataset showed reduced performance when applied "out of the box" to an Australian dataset. Local retraining with transfer learning using available model weights improved model performance. Keywords: Screening Mammography, Convolutional Neural Network (CNN), Deep Learning Algorithms, Breast Cancer Supplemental material is available for this article. © RSNA, 2024 See also commentary by Cadrin-Chênevert in this issue.