Lessons Learned in Building Expertly Annotated Multi-Institution Datasets and Hosting the RSNA AI Challenges.

The Radiological Society of North America (RSNA) has held artificial intelligence competitions to tackle real-world medical imaging problems at least annually since 2017. This article examines the challenges and processes involved in organizing these competitions, with a specific emphasis on the creation and curation of high-quality datasets. The collection of diverse and representative medical imaging data involves dealing with issues of patient privacy and data security. Furthermore, ensuring quality and consistency in data, which includes expert labeling and accounting for various patient and imaging characteristics, necessitates substantial planning and resources. Overcoming these obstacles requires meticulous project management and adherence to strict timelines. The article also highlights the potential of crowdsourced annotation to progress medical imaging research. Through the RSNA competitions, an effective global engagement has been realized, resulting in innovative solutions to complex medical imaging problems, thus potentially transforming health care by enhancing diagnostic accuracy and patient outcomes. Keywords: Use of AI in Education, Artificial Intelligence © RSNA, 2024.

Performance of the Winning Algorithms of the RSNA 2022 Cervical Spine Fracture Detection Challenge.

Purpose To evaluate and report the performance of the winning algorithms of the Radiological Society of North America Cervical Spine Fracture AI Challenge. Materials and Methods The competition was open to the public on Kaggle from July 28 to October 27, 2022. A sample of 3112 CT scans with and without cervical spine fractures (CSFx) were assembled from multiple sites (12 institutions across six continents) and prepared for the competition. The test set had 1093 scans (private test set: n = 789; mean age, 53.40 years ± 22.86 [SD]; 509 males; public test set: n = 304; mean age, 52.51 years ± 20.73; 189 males) and 847 fractures. The eight top-performing artificial intelligence (AI) algorithms were retrospectively evaluated, and the area under the receiver operating characteristic curve (AUC) value, F1 score, sensitivity, and specificity were calculated. Results A total of 1108 contestants composing 883 teams worldwide participated in the competition. The top eight AI models showed high performance, with a mean AUC value of 0.96 (95% CI: 0.95, 0.96), mean F1 score of 90% (95% CI: 90%, 91%), mean sensitivity of 88% (95% Cl: 86%, 90%), and mean specificity of 94% (95% CI: 93%, 96%). The highest values reported for previous models were an AUC of 0.85, F1 score of 81%, sensitivity of 76%, and specificity of 97%. Conclusion The competition successfully facilitated the development of AI models that could detect and localize CSFx on CT scans with high performance outcomes, which appear to exceed known values of previously reported models. Further study is needed to evaluate the generalizability of these models in a clinical environment. Keywords: Cervical Spine, Fracture Detection, Machine Learning, Artificial Intelligence Algorithms, CT, Head/Neck Supplemental material is available for this article. © RSNA, 2024.

The RSNA Cervical Spine Fracture CT Dataset.

This dataset is composed of cervical spine CT images with annotations related to fractures; it is available at https://www.kaggle.com/competitions/rsna-2022-cervical-spine-fracture-detection/.

Pearls & Oy-sters: ATX-FGF14 Mimicking Autoimmune Pathology.

ATX-FGF14 (formerly spinocerebellar ataxia 27, OMIM #193003) is an autosomal dominant condition caused by a pathogenic variant in the fibroblast growth factor 14 (FGF14, OMIM #601515) gene located on chromosome 13. The phenotypic expression can vary in patients with the same genotype, often delaying diagnosis, especially in probands without known affected relatives and/or with limited available family history. We describe 2 cases of ATX-FGF14 in 1 family with a focus on the importance of differentiating episodic manifestations of neurogenetic conditions from inflammatory/autoimmune neurologic conditions. A 68-year-old male patient (case 1) presented with episodic dysarthria, dizziness, imbalance, and encephalopathy, creating suspicion for a possible autoimmune etiology. At the first evaluation, the patient reported no significant family history. Four years later, on revisiting the family history, he noted that his 49-year-old niece (case 2) had also developed neurologic symptoms of an unclear etiology. On evaluation, she had tremor and ataxia. Both patients also had coexistent evidence of systemic autoimmunity that likely contributed to the initial suspicion of neurologic autoimmunity, and neither had cerebellar or brainstem volume loss. Ultimately, their genetic testing revealed a pathogenic structural variant in the FGF14 gene, consistent with ATX-FGF14. These 2 cases highlight the importance of a detailed interval family history at each visit, especially in undiagnosed adult patients, as well as the importance of objectively analyzing the impact of immunotherapy diagnostic treatment trials to avoid unnecessary immunomodulatory medications.

The arterial input function: Spatial dependence within the imaging volume and its influence on 3D quantitative dynamic contrast-enhanced MRI for head and neck cancer.

PURPOSE: To evaluate the dependence of the arterial input function (AIF) on the imaging z-axis and its effect on 3D DCE MRI pharmacokinetic parameters as mediated by the SPGR signal equation and Extended Tofts-Kermode model. THEORY: For SPGR-based 3D DCE MRI acquisition of the head and neck, inflow effects within vessels violate the assumptions underlying the SPGR signal model. Errors in the SPGR-based AIF estimate propagate through the Extended Tofts-Kermode model to affect the output pharmacokinetic parameters. MATERIALS AND METHODS: 3D DCE-MRI data were acquired for six newly diagnosed HNC patients in a prospective single arm cohort study. AIF were selected within the carotid arteries at each z-axis location. A region of interest (ROI) was placed in normal paravertebral muscle and the Extended Tofts-Kermode model solved for each pixel within the ROI for each AIF. Results were compared to those obtained with a published population average AIF. RESULTS: Due to inflow effect, the AIF showed extreme variation in their temporal shapes. Ktrans was most sensitive to the initial bolus concentration and showed more variation over the muscle ROI with AIF taken from the upstream portion of the carotid. kep was less sensitive to the peak bolus concentration and showed less variation for AIF taken from the upstream portion of the carotid. CONCLUSION: Inflow effects may introduce an unknown bias to SPGR-based 3D DCE pharmacokinetic parameters. Variation in the computed parameters depends on the selected AIF location. In the context of high flow, measurements may be limited to relative rather than absolute quantitative parameters.

Application of Radiomics to the Differential Diagnosis of Temporal Bone Skull Base Lesions: A Pilot Study.

BACKGROUND: Temporal bone skull base pathologies represent a complex differential because they can be radiographically obscure and difficult to diagnose without biopsy. Radiomics involves the use of mathematical quantification of imaging data beyond simple intensity, size, and location to inform diagnosis and prognosis. We examined the feasibility of using radiomic parameters to help predict temporal bone tumor type. METHODS: A total of 117 radiomic parameters were analyzed from 5 magnetic resonance imaging sequences (T1 without contrast, T1 with contrast, T2, fluid-attenuated inversion recovery, apparent diffusion coefficient [ADC]) for each tumor. Statistical analysis was used to delineate known primary, metastatic/secondary, and lymphoma lesions using radiomics. RESULTS: The mean tumor volumes for the 14 primary, 12 secondary, and 8 lymphoma lesions were 2.98 ± 2.11, 3.28 ± 2.31, and 12.16 ± 7.1 cm3, respectively (P = 0.2). No significant differences in mean intensity values for any sequence helped distinguish tumors (P > 0.05), but 6 radiomic parameters were significantly correlated with diagnostic accuracy. Discriminant analysis using a stepwise algorithm generated a model where radiomic parameters for T1 cluster prominence, ADC dependence nonuniformity, T1 with contrast zone percentage, and ADC informational measure of correlation 2 achieved the best predictive model (P = 0.0001). These significant characteristics were often indirect measures of tumor heterogeneity on different magnetic resonance imaging sequences. CONCLUSIONS: These data suggest that quantitative measures of tumor heterogeneity can be discriminatory of pathology and might be integrated into clinical workflow. Although this pilot study requires further validation, these data support the exploration of radiomics in temporal bone radiographic diagnostics.

Deep Learning Subtraction Angiography: Improved Generalizability with Transfer Learning.

PURPOSE: To investigate the utility and generalizability of deep learning subtraction angiography (DLSA) for generating synthetic digital subtraction angiography (DSA) images without misalignment artifacts. MATERIALS AND METHODS: DSA images and native digital angiograms of the cerebral, hepatic, and splenic vasculature, both with and without motion artifacts, were retrospectively collected. Images were divided into a motion-free training set (n = 66 patients, 9,161 images) and a motion artifact-containing test set (n = 22 patients, 3,322 images). Using the motion-free set, the deep neural network pix2pix was trained to produce synthetic DSA images without misalignment artifacts directly from native digital angiograms. After training, the algorithm was tested on digital angiograms of hepatic and splenic vasculature with substantial motion. Four board-certified radiologists evaluated performance via visual assessment using a 5-grade Likert scale. Subgroup analyses were performed to analyze the impact of transfer learning and generalizability to novel vasculature. RESULTS: Compared with the traditional DSA method, the proposed approach was found to generate synthetic DSA images with significantly fewer background artifacts (a mean rating of 1.9 [95% CI, 1.1-2.6] vs 3.5 [3.5-4.4]; P = .01) without a significant difference in foreground vascular detail (mean rating of 3.1 [2.6-3.5] vs 3.3 [2.8-3.8], P = .19) in both the hepatic and splenic vasculature. Transfer learning significantly improved the quality of generated images (P < .001). CONCLUSIONS: DLSA successfully generates synthetic angiograms without misalignment artifacts, is improved through transfer learning, and generalizes reliably to novel vasculature that was not included in the training data.

"Fully automated segmentation of the corticospinal tract using the TractSeg algorithm in patients with brain tumors".

OBJECTIVE: Tractography has been used to define the presurgical location of white matter tracts, but this is subjective and time-intensive, making incorporation to imaging workflow at scale problematic. The objective is to validate a fully automated pipeline using the TractSeg algorithm (Wasserthal et al. NeuroImage 2018;183:239-253) to segment the corticospinal tract in patients with brain tumors adjacent to the corticospinal tract. METHODS: The process of importing a structural MPRAGE sequence and raw diffusion weighted images from PACS, executing the TractSeg algorithm, overlaying the resulting bilateral corticospinal tracts on the MPRAGE image, and exporting this composite image to PACS was automated. This procedure was used to segment the corticospinal tract in 28 patients with brain masses adjacent to or displacing the corticospinal tract. These segmentations were compared with both manual deterministic tractography performed with DSI Studio using seeds placed in the pons and an automated tractography method in DSI Studio. RESULTS: The automated algorithm was able to segment the bilateral corticospinal tracts in all 28 patients whereas the manual reference method and DSI Studio based automated tractography were unsuccessful in 2 and 1 patients, respectively. In all cases, the TractSeg segmentations very closely matched the manual segmentations. Also, TractSeg appeared to include larger portions of the lateral corticospinal tract fibers than the other 2 methods. CONCLUSION: The TractSeg algorithm demonstrated robust performance in segmenting the corticospinal tract in patients with brain tumors adjacent to this tract. The algorithm is fast to perform and has great potential for optimizing and streamlining neurosurgical planning.

Radiology Performed Fluoroscopy-Guided Lumbar Punctures Decrease Volume of Diagnostic Study Interpretation - Impact on Resident Training and Potential Solutions.

OBJECTIVES: Lumbar punctures performed in radiology departments have significantly increased over the last few decades and are typically performed in academic centers by radiology trainees using fluoroscopy guidance. Performing fluoroscopy-guided lumbar punctures (FGLPs) can often constitute a large portion of a trainee's workday and the impact of performing FGLPs on the trainee's clinical productivity (i.e. dictating reports on neuroradiology cross-sectional imaging) has not been studied. The purpose of the study was to evaluate the relationship between the number of FGLPs performed and cross-sectional neuroimaging studies dictated by residents during their neuroradiology rotation (NR). MATERIAL AND METHODS: The number of FGLPs and myelograms performed and neuroimaging studies dictated by radiology residents on our neuroradiology service from July 2008 to December 2017 were retrospectively reviewed. The relationship between the number of FGLPs performed and neuroimaging studies (CT and MRI) dictated per day by residents was examined. RESULTS: Radiology residents (n = 84) performed 3437 FGLPs and myelograms and interpreted 33402 cross-sectional studies. Poisson regression demonstrated an exponential decrease in number of studies dictated daily with a rising number of FGLPs performed (P = 0.0001) and the following formula was derived: Number of expected studies dictated per day assuming no FGLPs × e-0.25 x number of FGLPs = adjusted expected studies dictated for the day. CONCLUSION: We quantified the impact performing FGLPs can have on the number of neuroimaging reports residents dictate on the NR. We described solutions to potentially decrease unnecessary FGLP referrals including establishing departmental guidelines for FGLP referrals and encouraging bedside lumbar punctures attempts before referral. We also emphasized equally distributing the FGLPs among trainees to mitigate procedural burden.

Systemic Therapies for Melanoma Brain Metastases: A Primer for Radiologists.

OBJECTIVE: The purpose of this article is to provide a primer for radiologists outlining the modern systemic therapies used in melanoma brain metastases, including tyrosine kinase inhibitors and immune checkpoint inhibitors. The role of radiologic treatment response evaluation will be discussed from the standpoint of both modern systemic therapies and more traditional treatments. CONCLUSION: Understanding the role of systemic treatments in melanoma brain metastases is critical for oncologic imaging interpretation in this unique patient population.

Visceral Adiposity Predicts Severity of Acute Pancreatitis.

OBJECTIVES: The aim of this study was to determine the association of visceral adiposity with severe outcomes in acute pancreatitis (AP). METHODS: This retrospective study included consecutive patients with AP admitted to a tertiary care hospital between January 2010 and January 2015 who underwent a computed tomography scan. The visceral adipose tissue (VAT) volume was estimated using the method of Linder and colleagues. Multivariable logistic regression analysis was conducted to assess VAT as a predictor of severe AP compared with other validated predictors of severity. RESULTS: Five hundred and seventy four patients were admitted during the study period, of which 252 had a computed tomography scan available. Patients with severe AP had a larger VAT area compared with those with mild or moderate AP (mean: 184.9 cm vs 79.9 cm, P = 0.006). Patients who developed multisystem organ failure or had acute necrotic collections had a larger VAT area than those who did not (150.6 cm vs 91.0 cm, P = 0.004 and 174.0 cm vs 91.9 cm, P = 0.003, respectively). Visceral adipose tissue area demonstrated superior discrimination of severe AP compared with other severity predictors. CONCLUSIONS: Increased VAT area is a strong predictor of severe pancreatitis, necrosis, and multisystem organ failure.

Making classifying selectors work for foam elimination in the activated-sludge process.

Classifying selectors are used to control the population of foam-causing organisms in activated-sludge plants to prevent the development of nuisance foams. The term, classifying selector, refers to the physical mechanism by which these organisms are selected against; foam-causing organisms are enriched into the solids in the foam and their rapid removal controls their population at low levels in the mixed liquor. Foam-causing organisms are wasted "first" rather than accumulating on the surface of tanks and thereby being wasted "last", which is typical of the process. This concept originated in South Africa, where pilot studies showed that placement of a flotation tank for foam removal prior to secondary clarifiers would eliminate foam-causing organisms. It was later simplified in the United States by using the aeration in aeration tanks or aerated channels coupled with simple baffling and adjustable weirs to make continuous separation of nuisance organisms from the mixed liquor.