Toward Precision Diagnosis: Machine Learning in Identifying Malignant Orbital Tumors With Multiparametric 3 T MRI.

BACKGROUND: Orbital tumors present a diagnostic challenge due to their varied locations and histopathological differences. Although recent advancements in imaging have improved diagnosis, classification remains a challenge. The integration of artificial intelligence in radiology and ophthalmology has demonstrated promising outcomes. PURPOSE: This study aimed to evaluate the performance of machine learning models in accurately distinguishing malignant orbital tumors from benign ones using multiparametric 3 T magnetic resonance imaging (MRI) data. MATERIALS AND METHODS: In this single-center prospective study, patients with orbital masses underwent presurgery 3 T MRI scans between December 2015 and May 2021. The MRI protocol comprised multiparametric imaging including dynamic contrast-enhanced (DCE), diffusion-weighted imaging (DWI), intravoxel incoherent motion (IVIM), as well as morphological imaging acquisitions. A repeated nested cross-validation strategy using random forest classifiers was used for model training and evaluation, considering 8 combinations of explanatory features. Shapley additive explanations (SHAP) values were used to assess feature contributions, and the model performance was evaluated using multiple metrics. RESULTS: One hundred thirteen patients were analyzed (57/113 [50.4%] were women; average age was 51.5 ± 17.5 years, range: 19-88 years). Among the 8 combinations of explanatory features assessed, the performance on predicting malignancy when using the most comprehensive model, which is the most exhaustive one incorporating all 46 explanatory features-including morphology, DWI, DCE, and IVIM, achieved an area under the curve of 0.9 [0.73-0.99]. When using the streamlined "10-feature signature" model, performance reached an area under the curve of 0.88 [0.71-0.99]. Random forest feature importance graphs measured by the mean of SHAP values pinpointed the 10 most impactful features, which comprised 3 quantitative IVIM features, 4 quantitative DCE features, 1 quantitative DWI feature, 1 qualitative DWI feature, and age. CONCLUSIONS: Our findings demonstrate that a machine learning approach, integrating multiparametric MRI data such as DCE, DWI, IVIM, and morphological imaging, offers high-performing models for differentiating malignant from benign orbital tumors. The streamlined 10-feature signature, with a performance close to the comprehensive model, may be more suitable for clinical application.

Diagnostic Performance of Dynamic Contrast-Enhanced 3T MR Imaging for Characterization of Orbital Lesions: Validation in a Large Prospective Study.

BACKGROUND AND PURPOSE: Orbital lesions are rare but serious. Their characterization remains challenging. Diagnosis is based on biopsy or surgery, which implies functional risks. It is necessary to develop noninvasive diagnostic tools. The goal of this study was to evaluate the diagnostic performance of dynamic contrast-enhanced MR imaging at 3T when distinguishing malignant from benign orbital tumors on a large prospective cohort. MATERIALS AND METHODS: This institutional review board-approved prospective single-center study enrolled participants presenting with an orbital lesion undergoing a 3T MR imaging before surgery from December 2015 to May 2021. Morphologic, diffusion-weighted, and dynamic contrast-enhanced MR images were assessed by 2 readers blinded to all data. Univariable and multivariable analyses were performed. To assess diagnostic performance, we used the following metrics: area under the curve, sensitivity, and specificity. Histologic analysis, obtained through biopsy or surgery, served as the criterion standard for determining the benign or malignant status of the tumor. RESULTS: One hundred thirty-one subjects (66/131 [50%] women and 65/131 [50%] men; mean age, 52 [SD, 17.1] years; range, 19-88 years) were enrolled. Ninety of 131 (69%) had a benign lesion, and 41/131 (31%) had a malignant lesion. Univariable analysis showed a higher median of transfer constant from blood plasma to the interstitial environment (K trans) and of transfer constant from the interstitial environment to the blood plasma (minute-1) (Kep) and a higher interquartile range of K trans in malignant-versus-benign lesions (1.1 minute-1 versus 0.65 minute-1, P = .03; 2.1 minute-1 versus 1.1 minute-1, P = .01; 0.81 minute-1 versus 0.65 minute-1, P = .009, respectively). The best-performing multivariable model in distinguishing malignant-versus-benign lesions included parameters from dynamic contrast-enhanced imaging, ADC, and morphology and reached an area under the curve of 0.81 (95% CI, 0.67-0.96), a sensitivity of 0.82 (95% CI, 0.55-1), and a specificity of 0.81 (95% CI, 0.65-0.96). CONCLUSIONS: Dynamic contrast-enhanced MR imaging at 3T appears valuable when characterizing orbital lesions and provides complementary information to morphologic imaging and DWI.

Best imaging signs identified by radiomics could outperform the model: application to differentiating lung carcinoid tumors from atypical hamartomas.

OBJECTIVES: Lung carcinoids and atypical hamartomas may be difficult to differentiate but require different treatment. The aim was to differentiate these tumors using contrast-enhanced CT semantic and radiomics criteria. METHODS: Between November 2009 and June 2020, consecutives patient operated for hamartomas or carcinoids with contrast-enhanced chest-CT were retrospectively reviewed. Semantic criteria were recorded and radiomics features were extracted from 3D segmentations using Pyradiomics. Reproducible and non-redundant radiomics features were used to training a random forest algorithm with cross-validation. A validation-set from another institution was used to evaluate of the radiomics signature, the 3D 'median' attenuation feature (3D-median) alone and the mean value from 2D-ROIs. RESULTS: Seventy-three patients (median 58 years [43‒70]) were analyzed (16 hamartomas; 57 carcinoids). The radiomics signature predicted hamartomas vs carcinoids on the external dataset (22 hamartomas; 32 carcinoids) with an AUC = 0.76. The 3D-median was the most important in the model. Density thresholds < 10 HU to predict hamartoma and > 60 HU to predict carcinoids were chosen for their high specificity > 0.90. On the external dataset, sensitivity and specificity of the 3D-median and 2D-ROIs were, respectively, 0.23, 1.00 and 0.13, 1.00 < 10 HU; 0.63, 0.95 and 0.69, 0.91 > 60 HU. The 3D-median was more reproducible than 2D-ROIs (ICC = 0.97 95% CI [0.95‒0.99]; bias: 3 ± 7 HU limits of agreement (LoA) [- 10‒16] vs. ICC = 0.90 95% CI [0.85‒0.94]; bias: - 0.7 ± 21 HU LoA [- 4‒40], respectively). CONCLUSIONS: A radiomics signature can distinguish hamartomas from carcinoids with an AUC = 0.76. Median density < 10 HU and > 60 HU on 3D or 2D-ROIs may be useful in clinical practice to diagnose these tumors with confidence, but 3D is more reproducible. CRITICAL RELEVANCE STATEMENT: Radiomic features help to identify the most discriminating imaging signs using random forest. 'Median' attenuation value (Hounsfield units), extracted from 3D-segmentations on contrast-enhanced chest-CTs, could distinguish carcinoids from atypical hamartomas (AUC = 0.85), was reproducible (ICC = 0.97), and generalized to an external dataset. KEY POINTS: • 3D-'Median' was the best feature to differentiate carcinoids from atypical hamartomas (AUC = 0.85). • 3D-'Median' feature is reproducible (ICC = 0.97) and was generalized to an external dataset. • Radiomics signature from 3D-segmentations differentiated carcinoids from atypical hamartomas with an AUC = 0.76. • 2D-ROI value reached similar performance to 3D-'median' but was less reproducible (ICC = 0.90).

Comparative performances of machine learning algorithms in radiomics and impacting factors.

There are no current recommendations on which machine learning (ML) algorithms should be used in radiomics. The objective was to compare performances of ML algorithms in radiomics when applied to different clinical questions to determine whether some strategies could give the best and most stable performances regardless of datasets. This study compares the performances of nine feature selection algorithms combined with fourteen binary classification algorithms on ten datasets. These datasets included radiomics features and clinical diagnosis for binary clinical classifications including COVID-19 pneumonia or sarcopenia on CT, head and neck, orbital or uterine lesions on MRI. For each dataset, a train-test split was created. Each of the 126 (9 × 14) combinations of feature selection algorithms and classification algorithms was trained and tuned using a ten-fold cross validation, then AUC was computed. This procedure was repeated three times per dataset. Best overall performances were obtained with JMI and JMIM as feature selection algorithms and random forest and linear regression models as classification algorithms. The choice of the classification algorithm was the factor explaining most of the performance variation (10% of total variance). The choice of the feature selection algorithm explained only 2% of variation, while the train-test split explained 9%.

Co-registration with subtraction and color-coding or fusion improves the detection of new and growing lesions on follow-up MRI examination of patients with multiple sclerosis.

PURPOSE: The purpose of this study was to compare the performance of three magnetic resonance imaging (MRI) reading methods in the follow-up of patients with multiple sclerosis (MS). MATERIALS AND METHODS: This retrospective study included patients with MS who underwent two brain follow-up MRI examinations with three-dimensional fluid-attenuated inversion recovery (FLAIR) sequences between September 2016 and December 2019. Two neuroradiology residents independently reviewed FLAIR images using three post-processing methods including conventional reading (CR), co-registration fusion (CF), and co-registration subtraction with color-coding (CS), while being blinded to all data but FLAIR images. The presence and number of new, growing, or shrinking lesions were compared between reading methods. The reading time, reading confidence, and inter- and intra-observer agreements were also assessed. An expert neuroradiologist established the standard of reference. Statistical analyses were corrected for multiple testing. RESULTS: A total of 198 patients with MS were included. There were 130 women and 68 men, with a mean age of 41 ± 12 (standard deviation) years (age range: 21-79 years). Using CS and CF, more patients were detected with new lesions compared to CR (93/198 [47%] and 79/198 [40%] vs. 54/198 [27%], respectively; P < 0.01). The median number of new hyperintense FLAIR lesions detected was significantly greater using CS and CF compared to CR (2 [Q1, Q3: 0, 6] and 1 [Q1, Q3: 0, 3] vs. 0 [Q1, Q3: 0, 1], respectively; P < 0.001). The mean reading time was significantly shorter using CS and CF compared to CR (P < 0.001), with higher confidence in readings and higher inter- and intra-observer agreements. CONCLUSION: Post-processing tools such as CS and CF substantially improve the accuracy of follow-up MRI examinations in patients with MS while reducing reading time and increasing readers' confidence and reproducibility.

Revolutionizing radiology with GPT-based models: Current applications, future possibilities and limitations of ChatGPT.

Artificial intelligence has demonstrated utility and is increasingly being used in the field of radiology. The use of generative pre-trained transformer (GPT)-based models has the potential to revolutionize the field of radiology, offering new possibilities for improving accuracy, efficiency, and patient outcome. Current applications of GPT-based models in radiology include report generation, educational support, clinical decision support, patient communication, and data analysis. As these models continue to advance and improve, it is likely that more innovative uses for GPT-based models in the field of radiology at large will be developed, further enhancing the role of technology in the diagnostic process. ChatGPT is a variant of GPT that is specifically fine-tuned for conversational language understanding and generation. This article reports some answers provided by ChatGPT to various questions that radiologists may have regarding ChatGPT and identifies the potential benefits ChatGPT may offer in their daily practice but also current limitations. Similar to other applications of artificial intelligence in the field of imaging, further formal validation of ChatGPT is required.

Contrast-Enhanced 3D Spin Echo T1-Weighted Sequence Outperforms 3D Gradient Echo T1-Weighted Sequence for the Detection of Multiple Sclerosis Lesions on 3.0 T Brain MRI.

BACKGROUND: Using reliable contrast-enhanced T1 sequences is crucial to detect enhancing brain lesions for multiple sclerosis (MS) at the time of diagnosis and over follow-up. Contrast-enhanced 3D gradient-recalled echo (GRE) T1-weighted imaging (WI) and 3D turbo spin echo (TSE) T1-WI are both available for clinical practice and have never been compared within the context of this diagnosis. PURPOSE: The aim of this study was to compare contrast-enhanced 3D GRE T1-WI and 3D TSE T1-WI for the detection of enhancing lesions in the brains of MS patients. METHODS: This single-center prospective study enrolled patients with MS who underwent a 3.0 T brain MRI from August 2017 to April 2021 for follow-up. Contrast-enhanced 3D GRE T1-WI and 3D TSE T1-WI were acquired in randomized order. Two independent radiologists blinded to all data reported all contrast-enhanced lesions in each sequence. Their readings were compared with a reference standard established by a third expert neuroradiologist. Interobserver agreement, contrast ratio, and contrast-to-noise ratio were calculated for both sequences. RESULTS: A total of 158 MS patients were included (mean age, 40 ± 11 years; 95 women). Significantly more patients had at least 1 contrast-enhanced lesion on 3D TSE T1-WI than on 3D GRE T1-WI for both readers (61/158 [38.6%] vs 48/158 [30.4%] and 60/158 [38.6%] vs 47/158 [29.7%], P < 0.001). Significantly more contrast-enhanced lesions per patient were detected on 3D TSE T1-WI (mean 2.47 vs 1.56 and 2.56 vs 1.39, respectively, P < 0.001). Interobserver agreement was excellent for both sequences, κ = 0.96 (confidence interval [CI], 0.91-1.00) for 3D TSE T1-WI and 0.92 (CI, 0.86-0.99) for 3D GRE T1-WI. Contrast ratio and contrast-to-noise ratio were significantly higher on 3D TSE T1-WI (0.84 vs 0.53, P < 0.001, and 87.9 vs 57.8, P = 0.03, respectively). CONCLUSIONS: At 3.0 T, contrast-enhanced 3D TSE-T1-WI supports the detection of significantly more enhancing lesions than 3D GRE T1-WI and should therefore be used for MS patients requiring contrast-enhanced examination.

Artificial intelligence in diagnostic and interventional radiology: Where are we now?

The emergence of massively parallel yet affordable computing devices has been a game changer for research in the field of artificial intelligence (AI). In addition, dramatic investment from the web giants has fostered the development of a high-quality software stack. Going forward, the combination of faster computers with dedicated software libraries and the widespread availability of data has opened the door to more flexibility in the design of AI models. Radiomics is a process used to discover new imaging biomarkers that has multiple applications in radiology and can be used in conjunction with AI. AI can be used throughout the various processes of diagnostic imaging, including data acquisition, reconstruction, analysis and reporting. Today, the concept of "AI-augmented" radiologists is preferred to the theory of the replacement of radiologists by AI in many indications. Current evidence bolsters the assumption that AI-assisted radiologists work better and faster. Interventional radiology becomes a data-rich specialty where the entire procedure is fully recorded in a standardized DICOM format and accessible via standard picture archiving and communication systems. No other interventional specialty can bolster such readiness. In this setting, interventional radiology could lead the development of AI-powered applications in the broader interventional community. This article provides an update on the current status of radiomics and AI research, analyzes upcoming challenges and also discusses the main applications in AI in interventional radiology to help radiologists better understand and criticize articles reporting AI in medical imaging.

A signature of structural MRI features at 3 Tesla allows an accurate characterization of orbital cavernous venous malformation.

OBJECTIVES: To differentiate OCVM from other orbital lesions using structural MRI. METHODS: This IRB-approved a historical-prospective cohort single-center analysis of a prospective cohort that included consecutive adult patients presenting with an orbital lesion undergoing a 3T MRI before surgery from December 2015 to May 2021. Two readers blinded to all data read all MRIs assessing structural MRI characteristics. A univariate analysis followed by a stepwise multivariate analysis identified structural MRI features showing the highest sensitivity and specificity when diagnosing OCVM. RESULTS: One hundred ninety-one patients with 30/191 (16%) OCVM and 161/191 (84%) other orbital lesions were included. OCVM were significantly more likely to present with a higher signal intensity than that of the cortex on T2WI: 26/29 (89.7%) versus 28/160 (17.5%), p < 0.001, or with a chemical shift artifact (CSA): 26/29 (89.7%) versus 16/155 (10.3%), p < 0.001, or to present with a single starting point of enhancement, as compared to other orbital lesions: 18/29 (62.1%) versus 4/159 (2.5%), p = 0.001. The step-wise analysis identified 2 signatures increasing performances. Signature 1 combined a higher signal intensity than that of the cortex on T2WI and a CSA. Signature 2 included these two features and the presence of a single starting point of enhancement. Sensitivity, specificity, and accuracy were 0.83, 0.94, and 0.92 for signature 1 and 0.97, 0.93, and 0.93 for signature 2, respectively. CONCLUSION: Structural MRI yields high sensitivity and specificity when diagnosing OCVM. KEY POINTS: • Structural MRI shows high sensitivity and specificity when diagnosing orbital cavernous venous malformation. • We identified two signatures combining structural MRI features which might be used easily in routine clinical practice. • The combination of higher signal intensity of the lesion as compared to the cortex on T2WI and of a chemical shift artifact yields a sensitivity and specificity of 0.83 and 0.94 for the diagnosis of orbital cavernous venous malformation, respectively.

Comparison between 7 Tesla and 3 Tesla MRI for characterizing orbital lesions.

PURPOSE: Characterizing orbital lesions remains challenging with imaging. The purpose of this study was to compare 3 Tesla (T) to 7 T magnetic resonance imaging (MRI) for characterizing orbital lesions. MATERIALS AND METHODS: This prospective single-center study enrolled participants presenting with orbital lesions from May to October 2019, who underwent both 7 T and 3 T MRI examinations. Two neuroradiologists, blinded to all data, read both datasets independently and randomly. They assessed general characteristics of each orbital lesion as well as image quality and presence of artifacts. Comparison between both datasets was made using Fisher exact test. RESULTS: Seven patients (4 women, 3 men) with a median age of 52 years were enrolled. Orbital lesion conspicuity was better scored at 7 T compared to 3 T MRI, with 3/7 lesions (43%) scored as very conspicuous at 7 T compared to 0/7 lesion (0%) at 3 T, although the difference was not significant (P = 0.16). Delineation of lesion margins was better scored at 7 T compared to 3 T with 3/7 lesions (43%) scored as very well delineated on 7 T compared to 0/7 lesions (0%) at 3 T, although the difference was not significant (P = 0.34). Details of internal structure were better assessed at 7 T compared to 3 T, with 4/7 lesions (57%) displaying numerous internal details compared to 0/7 lesions (0%) at 3 T (P = 0.10). Internal microvessels were visible in 3/7 lesions (43%) at 7 T compared to 0/7 lesions (0%) at 3 T (P = 0.19). CONCLUSION: Although no significant differences were found between 7 T and 3 T MRI, assumably due to a limited number of patients, our study suggests that 7 Tesla MRI might help improve the characterization of orbital lesions. However, further studies with more patients are needed.

Validation of a deep learning segmentation algorithm to quantify the skeletal muscle index and sarcopenia in metastatic renal carcinoma.

OBJECTIVES: To validate a deep learning (DL) algorithm for measurement of skeletal muscular index (SMI) and prediction of overall survival in oncology populations. METHODS: A retrospective single-center observational study included patients with metastatic renal cell carcinoma between 2007 and 2019. A set of 37 patients was used for technical validation of the algorithm, comparing manual vs DL-based evaluations. Segmentations were compared using mean Dice similarity coefficient (DSC), SMI using concordance correlation coefficient (CCC) and Bland-Altman plots. Overall survivals (OS) were compared using log-rank (Kaplan-Meier) and Mann-Whitney tests. Generalizability of the prognostic value was tested in an independent validation population (N = 87). RESULTS: Differences between two manual segmentations (DSC = 0.91, CCC = 0.98 for areas) or manual vs. automated segmentation (DSC = 0.90, CCC = 0.98 for areas, CCC = 0.97 for SMI) had the same order of magnitude. Bland-Altman plots showed a mean difference of -3.33 cm2 [95%CI: -15.98, 9.1] between two manual segmentations, and -3.28 cm2 [95% CI: -14.77, 8.21] for manual vs. automated segmentations. With each method, 20/37 (56%) patients were classified as sarcopenic. Sarcopenic vs. non-sarcopenic groups had statistically different survival curves with median OS of 6.0 vs. 12.5 (p = 0.008) and 6.0 vs. 13.9 (p = 0.014) months respectively for manual and DL methods. In the independent validation population, sarcopenic patients according to DL had a lower OS (10.7 vs. 17.3 months, p = 0.033). CONCLUSION: A DL algorithm allowed accurate estimation of SMI compared to manual reference standard. The DL-calculated SMI demonstrated a prognostic value in terms of OS. KEY POINTS: • A deep learning algorithm allows accurate estimation of skeletal muscle index compared to a manual reference standard with a concordance correlation coefficient of 0.97. • Sarcopenic patients according to SMI thresholds after segmentation by the deep learning algorithm had statistically significantly lower overall survival compared to non-sarcopenic patients.

Can we use radiomics in ultrasound imaging? Impact of preprocessing on feature repeatability.

PURPOSE: The purpose of this study was to assess the inter-slice radiomic feature repeatability in ultrasound imaging and the impact of preprocessing using intensity standardization and grey-level discretization to help improve radiomics reproducibility. MATERIALS AND METHODS: This single-center study enrolled consecutive patients with an orbital lesion who underwent ultrasound examination of the orbit from December 2015 to July 2019. Two images per lesion were randomly assigned to two subsets. Radiomic features were extracted and inter-slice repeatability was assessed using the intraclass correlation coefficient (ICC) between the subsets. The impact of preprocessing on feature repeatability was assessed using image intensity standardization with or without outliers removal on whole images, bounding boxes or regions of interest (ROI), and fixed bin size or fixed bin number grey-level discretization. Number of inter-slice repeatable features (ICC ≥0.7) between methods was compared. RESULTS: Eighty-eight patients (37 men, 51 women) with a mean age of 51.5 ± 17 (SD) years (range: 20-88 years) were enrolled. Without preprocessing, 29/101 features (28.7%) were repeatable between slices. The greatest number of repeatable features (41/101) was obtained using intensity standardization with outliers removal on the ROI and fixed bin size discretization. Standardization performed better with outliers removal than without (P < 0.001), and on ROIs than on native images (P < 0.001). Fixed bin size discretization performed better than fixed bin number (P = 0.008). CONCLUSION: Radiomic features extracted from ultrasound images are impacted by the slice and preprocessing. The use of intensity standardization with outliers removal applied to the ROI and a fixed bin size grey-level discretization may improve feature repeatability.

Digital health, big data and smart technologies for the care of patients with systemic autoimmune diseases: Where do we stand?

The past decade has seen tremendous development in digital health, including in innovative new technologies such as Electronic Health Records, telemedicine, virtual visits, wearable technology and sophisticated analytical tools such as artificial intelligence (AI) and machine learning for the deep-integration of big data. In the field of rare connective tissue diseases (rCTDs), these opportunities include increased access to scarce and remote expertise, improved patient monitoring, increased participation and therapeutic adherence, better patient outcomes and patient empowerment. In this review, we discuss opportunities and key-barriers to improve application of digital health technologies in the field of autoimmune diseases. We also describe what could be the fully digital pathway of rCTD patients. Smart technologies can be used to provide real-world evidence about the natural history of rCTDs, to determine real-life drug utilization, advanced efficacy and safety data for rare diseases and highlight significant unmet needs. Yet, digitalization remains one of the most challenging issues faced by rCTD patients, their physicians and healthcare systems. Digital health technologies offer enormous potential to improve autoimmune rCTD care but this potential has so far been largely unrealized due to those significant obstacles. The need for robust assessments of the efficacy, affordability and scalability of AI in the context of digital health is crucial to improve the care of patients with rare autoimmune diseases.

Assessment of an AI Aid in Detection of Adult Appendicular Skeletal Fractures by Emergency Physicians and Radiologists: A Multicenter Cross-sectional Diagnostic Study.

Background The interpretation of radiographs suffers from an ever-increasing workload in emergency and radiology departments, while missed fractures represent up to 80% of diagnostic errors in the emergency department. Purpose To assess the performance of an artificial intelligence (AI) system designed to aid radiologists and emergency physicians in the detection and localization of appendicular skeletal fractures. Materials and Methods The AI system was previously trained on 60 170 radiographs obtained in patients with trauma. The radiographs were randomly split into 70% training, 10% validation, and 20% test sets. Between 2016 and 2018, 600 adult patients in whom multiview radiographs had been obtained after a recent trauma, with or without one or more fractures of shoulder, arm, hand, pelvis, leg, and foot, were retrospectively included from 17 French medical centers. Radiographs with quality precluding human interpretation or containing only obvious fractures were excluded. Six radiologists and six emergency physicians were asked to detect and localize fractures with (n = 300) and fractures without (n = 300) the aid of software highlighting boxes around AI-detected fractures. Aided and unaided sensitivity, specificity, and reading times were compared by means of paired Student t tests after averaging of performances of each reader. Results A total of 600 patients (mean age ± standard deviation, 57 years ± 22; 358 women) were included. The AI aid improved the sensitivity of physicians by 8.7% (95% CI: 3.1, 14.2; P = .003 for superiority) and the specificity by 4.1% (95% CI: 0.5, 7.7; P < .001 for noninferiority) and reduced the average number of false-positive fractures per patient by 41.9% (95% CI: 12.8, 61.3; P = .02) in patients without fractures and the mean reading time by 15.0% (95% CI: -30.4, 3.8; P = .12). Finally, stand-alone performance of a newer release of the AI system was greater than that of all unaided readers, including skeletal expert radiologists, with an area under the receiver operating characteristic curve of 0.94 (95% CI: 0.92, 0.96). Conclusion The artificial intelligence aid provided a gain of sensitivity (8.7% increase) and specificity (4.1% increase) without loss of reading speed. © RSNA, 2021 Online supplemental material is available for this article.

Radiomic analysis of HTR-DCE MR sequences improves diagnostic performance compared to BI-RADS analysis of breast MR lesions.

PURPOSE: To assess the diagnostic performance of radiomic analysis using high temporal resolution (HTR)-dynamic contrast enhancement (DCE) MR sequences compared to BI-RADS analysis to distinguish benign from malignant breast lesions. MATERIALS AND METHODS: We retrospectively analyzed data from consecutive women who underwent breast MRI including HTR-DCE MR sequencing for abnormal enhancing lesions and who had subsequent pathological analysis at our tertiary center. Semi-quantitative enhancement parameters and textural features were extracted. Temporal change across each phase of textural features in HTR-DCE MR sequences was calculated and called "kinetic textural parameters." Statistical analysis by LASSO logistic regression and cross validation was performed to build a model. The diagnostic performance of the radiomic model was compared to the results of BI-RADS MR score analysis. RESULTS: We included 117 women with a mean age of 54 years (28-88). Of the 174 lesions analyzed, 75 were benign and 99 malignant. Seven semi-quantitative enhancement parameters and 57 textural features were extracted. Regression analysis selected 15 significant variables in a radiomic model (called "malignant probability score") which displayed an AUC = 0.876 (sensitivity = 0.98, specificity = 0.52, accuracy = 0.78). The performance of the malignant probability score to distinguish benign from malignant breast lesions (AUC = 0.876, 95%CI 0.825-0.925) was significantly better than that of BI-RADS analysis (AUC = 0.831, 95%CI 0.769-0.892). The radiomic model significantly reduced false positives (42%) with the same number of missed cancers (n = 2). CONCLUSION: A radiomic model including kinetic textural features extracted from an HTR-DCE MR sequence improves diagnostic performance over BI-RADS analysis. KEY POINTS: • Radiomic analysis using HTR-DCE is of better diagnostic performance (AUC = 0.876) than conventional breast MRI reading with BI-RADS (AUC = 0.831) (p < 0.001). • A radiomic malignant probability score under 19.5% gives a negative predictive value of 100% while a malignant probability score over 81% gives a positive predictive value of 100%. • Kinetic textural features extracted from HTR-DCE-MRI have a major role to play in distinguishing benign from malignant breast lesions.

A Magnetic Resonance Imaging Radiomics Signature to Distinguish Benign From Malignant Orbital Lesions.

OBJECTIVES: Distinguishing benign from malignant orbital lesions remains challenging both clinically and with imaging, leading to risky biopsies. The objective was to differentiate benign from malignant orbital lesions using radiomics on 3 T magnetic resonance imaging (MRI) examinations. MATERIALS AND METHODS: This institutional review board-approved prospective single-center study enrolled consecutive patients presenting with an orbital lesion undergoing a 3 T MRI prior to surgery from December 2015 to July 2019. Radiomics features were extracted from 6 MRI sequences (T1-weighted images [WIs], DIXON-T2-WI, diffusion-WI, postcontrast DIXON-T1-WI) using the Pyradiomics software. Features were selected based on their intraobserver and interobserver reproducibility, nonredundancy, and with a sequential step forward feature selection method. Selected features were used to train and optimize a Random Forest algorithm on the training set (75%) with 5-fold cross-validation. Performance metrics were computed on a held-out test set (25%) with bootstrap 95% confidence intervals (95% CIs). Five residents, 4 general radiologists, and 3 expert neuroradiologists were evaluated on their ability to visually distinguish benign from malignant lesions on the test set. Performance comparisons between reader groups and the model were performed using McNemar test. The impact of clinical and categorizable imaging data on algorithm performance was also assessed. RESULTS: A total of 200 patients (116 [58%] women and 84 [42%] men; mean age, 53.0 ± 17.9 years) with 126 of 200 (63%) benign and 74 of 200 (37%) malignant orbital lesions were included in the study. A total of 606 radiomics features were extracted. The best performing model on the training set was composed of 8 features including apparent diffusion coefficient mean value, maximum diameter on T1-WIs, and texture features. Area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity on the test set were respectively 0.869 (95% CI, 0.834-0.898), 0.840 (95% CI, 0.806-0.874), 0.684 (95% CI, 0.615-0.751), and 0.935 (95% CI, 0.905-0.961). The radiomics model outperformed all reader groups, including expert neuroradiologists (P < 0.01). Adding clinical and categorizable imaging data did not significantly impact the algorithm performance (P = 0.49). CONCLUSIONS: An MRI radiomics signature is helpful in differentiating benign from malignant orbital lesions and may outperform expert radiologists.

How can we combat multicenter variability in MR radiomics? Validation of a correction procedure.

OBJECTIVE: Test a practical realignment approach to compensate the technical variability of MR radiomic features. METHODS: T1 phantom images acquired on 2 scanners, FLAIR and contrast-enhanced T1-weighted (CE-T1w) images of 18 brain tumor patients scanned on both 1.5-T and 3-T scanners, and 36 T2-weighted (T2w) images of prostate cancer patients scanned in one of two centers were investigated. The ComBat procedure was used for harmonizing radiomic features. Differences in statistical distributions in feature values between 1.5- and 3-T images were tested before and after harmonization. The prostate studies were used to determine the impact of harmonization to distinguish between Gleason grades (GGs). RESULTS: In the phantom data, 40 out of 42 radiomic feature values were significantly different between the 2 scanners before harmonization and none after. In white matter regions, the statistical distributions of features were significantly different (p < 0.05) between the 1.5- and 3-T images for 37 out of 42 features in both FLAIR and CE-T1w images. After harmonization, no statistically significant differences were observed. In brain tumors, 41 (FLAIR) or 36 (CE-T1w) out of 42 features were significantly different between the 1.5- and 3-T images without harmonization, against 1 (FLAIR) or none (CE-T1w) with harmonization. In prostate studies, 636 radiomic features were significantly different between GGs after harmonization against 461 before. The ability to distinguish between GGs using radiomic features was increased after harmonization. CONCLUSION: ComBat harmonization efficiently removes inter-center technical inconsistencies in radiomic feature values and increases the sensitivity of studies using data from several scanners. KEY POINTS: • Radiomic feature values obtained using different MR scanners or imaging protocols can be harmonized by combining off-the-shelf image standardization and feature realignment procedures. • Harmonized radiomic features enable one to pool data from different scanners and centers without a substantial loss of statistical power caused by intra- and inter-center variability. • The proposed realignment method is applicable to radiomic features from different MR sequences and tumor types and does not rely on any phantom acquisition.