Accurate disease classification from radiology reports is essential for many applications. While supervised fine-tuning (SFT) of lightweight LLMs improves accuracy, it can degrade reasoning. We propose a two-stage approach: SFT on disease labels followed by Group Relative Policy Optimization (GRPO) to refine predictions by optimizing accuracy and format without reasoning supervision. Across three radiologist-annotated datasets, SFT outperformed baselines and GRPO further improved classification and enhanced reasoning recall and comprehensiveness.

Background: Dacryocystorhinostomy (DCR) is a surgery performed for nasolacrimal duct obstruction (NLDO). A high septal deflection may limit surgical access, and if not corrected, contribute to an increase in DCR failure rates due to increased adhesion formation. We hypothesize that a narrower distance between the nasal septum (NS) and lacrimal fossa (LF) is associated with higher rates of DCR failure. Objective: To evaluate whether a reduced nasal septum-lacrimal fossa (NS-LF) distance is associated with higher rates of endoscopic DCR (Endo DCR) failure and to determine a clinically relevant cutoff distance that may guide consideration for concomitant septoplasty. Methods: A single-institution retrospective chart review was performed for patients who received an endoscopic "Endo" DCR between July 2018 and July 2023. Data collected included demographics, septal deviation presence and laterality, NLDO laterality, and failure rates, where failure was defined as the need for a repeat ipsilateral DCR. The NS-LF distance was determined on preoperative computed tomography. Results: A total of 521 patients were identified, with an average age of 60.3 years (SD = 16.47); 4.4% (n = 23) of patients experienced Endo DCR failure, requiring revision Endo DCR; 8.7% (n = 45) of patients underwent concomitant septoplasty with initial DCR. Patients who failed had a significantly shorter NS-LF distance (8.47 mm vs 9.25 mm, P = .028). Receiver operating characteristic curve analysis cutoff distance of 8.85 mm demonstrated a NS-LF distance >= 8.85 mm had a significantly lower failure rate compared to those with a distance <8.85 mm (2.6% vs 7.3%, P = .016). All revision DCRs were successful regardless of septoplasty status. Conclusion: A NS-LF distance of <8.85 mm was significantly associated with Endo DCR failure, suggesting that concomitant septoplasty should be considered in these patients. Prospective studies and standardized preoperative imaging may improve surgical planning and outcomes in Endo DCR.

BACKGROUND AND PURPOSE:Normative diffusion values are essential for detecting microstructural abnormalities in the spinal cord. While diffusion kurtosis imaging (DKI) and neurite orientation dispersion and density imaging (NODDI) provide additional sensitivity beyond DTI, existing normative datasets are limited to adults with restricted cervical coverage. Pediatric spinal cord data remain scarce and predominantly DTI-based, leaving a gap in characterizing age-related microstructure. The purpose of this study was to establish a normative database of pediatric spinal cord diffusion metrics (DTI, DKI, NODDI) from C1-T12 using a clinically feasible Hybrid Diffusion Imaging (HYDI) protocol and to assess agreement between HYDI-derived and conventional DTI metrics. MATERIALS AND METHODS:One hundred typically developing children (mean age 12.07 ± 3.09 years), aged 6-17 years, underwent a 3T MRI scan including a T2-weighted, multi-shell, multiband, HYDI (b=0, 800, 2000 s/mm2) sequence covering C1-T12 and a validation subset, a single-shell DTI sequence covering C2-C5. Ten microstructural metrics were quantified for white matter (WM), gray matter (GM), and whole cord at each vertebral level. Multivariable regression modeled age and sex. Agreement between HYDI-derived and conventional DTI at C3 was evaluated using paired t-tests, Pearson correlation, and Bland-Altman analysis. RESULTS:Normative C1-T12 vertebral diffusion values were established. WM-GM contrast was evident for FA, NDI, MD, and AD, with significant differences across most cervical and thoracic segments. All metrics demonstrated significant age dependence (p <.01) after adjustment for sex, with higher FA, kurtosis, ODI and NDI and lower diffusivity and fISO in older children. Sex effects were modest and did not remain significant after false-discovery-rate correction. Accelerated HYDI demonstrated minimal bias versus conventional DTI at C3 with moderate individual-level agreement (r = 0.35-0.43; ICC 0.34-0.39). CONCLUSIONS:This study provides comprehensive normative reference values of pediatric spinal cord microstructure across C1-T12 and validates a time-efficient HYDI protocol for clinical translation.

The RSNA Lumbar Degenerative Imaging Spine Classification dataset is the largest publicly available adult MRI lumbar spine dataset for degenerative disease. The dataset includes multisequence, multiplanar MRIs from 2,697 patients through contributions from 8 institutions across 6 countries and 5 continents. ©RSNA, 2026.

Despite significant progress in applying large language models (LLMs) to the medical domain, several limitations still prevent them from practical applications. Among these are the constraints on model size and the lack of cohort-specific labeled datasets. In this work, we investigated the potential of improving a lightweight LLM, such as Llama 3.1-8B, through fine-tuning with datasets using synthetic labels. Two tasks are jointly trained by combining their respective instruction datasets. When the quality of the task-specific synthetic labels is relatively high (e.g., generated by GPT4-o), Llama 3.1-8B achieves satisfactory performance on the open-ended disease detection task, with a micro F1 score of 0.91. Conversely, when the quality of the task-relevant synthetic labels is relatively low (e.g., from the MIMIC-CXR dataset), fine-tuned Llama 3.1-8B is able to surpass its noisy teacher labels (micro F1 score of 0.67 v.s. 0.63) when calibrated against curated labels, indicating the strong inherent underlying capability of the model. These findings demonstrate the potential offine-tuning LLMs with synthetic labels, offering a promising direction for future research on LLM specialization in the medical domain.

Purpose To evaluate the performance of the winning machine learning models from the 2023 RSNA Abdominal Trauma Detection AI Challenge. Materials and Methods The competition was hosted on Kaggle and took place between July 26 and October 15, 2023. The multicenter competition dataset consisted of 4274 abdominal trauma CT scans, in which solid organs (liver, spleen, and kidneys) were annotated as healthy, low-grade, or high-grade injury. Studies were labeled as positive or negative for the presence of bowel and mesenteric injury and active extravasation. In this study, performances of the eight award-winning models were retrospectively assessed and compared using various metrics, including the area under the receiver operating characteristic curve (AUC), for each injury category. The reported mean values of these metrics were calculated by averaging the performance across all models for each specified injury type. Results The models exhibited strong performance in detecting solid organ injuries, particularly high-grade injuries. For binary detection of injuries, the models demonstrated mean AUC values of 0.92 (range, 0.90-0.94) for liver, 0.91 (range, 0.87-0.93) for splenic, and 0.94 (range, 0.93-0.95) for kidney injuries. The models achieved mean AUC values of 0.98 (range, 0.96-0.98) for high-grade liver, 0.98 (range, 0.97-0.99) for high-grade splenic, and 0.98 (range, 0.97-0.98) for high-grade kidney injuries. For the detection of bowel and mesenteric injuries and active extravasation, the models demonstrated mean AUC values of 0.85 (range, 0.74-0.93) and 0.85 (range, 0.79-0.89), respectively. Conclusion The award-winning models from the artificial intelligence challenge demonstrated strong performance in the detection of traumatic abdominal injuries on CT scans, particularly high-grade injuries. These models may serve as a performance baseline for future investigations and algorithms. Keywords: Abdominal Trauma, CT, American Association for the Surgery of Trauma, Machine Learning, Artificial Intelligence Supplemental material is available for this article. © RSNA, 2024.

BACKGROUND AND PURPOSE:Spinal cord injury (SCI) in the pediatric population presents a unique challenge in diagnosis and prognosis due to the complexity of performing clinical assessments on children. Accurate evaluation of structural changes in the spinal cord is essential for effective treatment planning. This study aims to evaluate structural characteristics in pediatric patients with SCI by comparing cross-sectional area (CSA), anterior-posterior (AP) width, and right-left (RL) width across all vertebral levels of the spinal cord between typically developing (TD) and participants with SCI. We employed deep learning techniques to utilize these measures for detecting SCI cases and determining their injury severity. MATERIALS AND METHODS:Sixty-one pediatric participants (ages 6-18), including 20 with chronic SCI and 41 TD, were enrolled and scanned by using a 3T MRI scanner. All SCI participants underwent the International Standards for Neurological Classification of Spinal Cord Injury (ISNCSCI) test to assess their neurologic function and determine their American Spinal Injury Association (ASIA) Impairment Scale (AIS) category. T2-weighted MRI scans were utilized to measure CSA, AP width, and RL widths along the entire cervical and thoracic cord. These measures were automatically extracted at every vertebral level of the spinal cord by using the spinal cord toolbox. Deep convolutional neural networks (CNNs) were utilized to classify participants into SCI or TD groups and determine their AIS classification based on structural parameters and demographic factors such as age and height. RESULTS:Significant differences (P < .05) were found in CSA, AP width, and RL width between SCI and TD participants, indicating notable structural alterations due to SCI. The CNN-based models demonstrated high performance, achieving 96.59% accuracy in distinguishing SCI from TD participants. Furthermore, the models determined AIS category classification with 94.92% accuracy. CONCLUSIONS:The study demonstrates the effectiveness of integrating cross-sectional structural imaging measures with deep learning methods for classification and severity assessment of pediatric SCI. The deep learning approach outperforms traditional machine learning models in diagnostic accuracy, offering potential improvements in patient care in pediatric SCI management.

Although there are relatively few diverse, high-quality medical imaging datasets on which to train computer vision artificial intelligence models, even fewer datasets contain expertly classified observations that can be repurposed to train or test such models. The traditional annotation process is laborious and time-consuming. Repurposing annotations and consolidating similar types of annotations from disparate sources has never been practical. Until recently, the use of natural language processing to convert a clinical radiology report into labels required custom training of a language model for each use case. Newer technologies such as large language models have made it possible to generate accurate and normalized labels at scale, using only clinical reports and specific prompt engineering. The combination of automatically generated labels extracted and normalized from reports in conjunction with foundational image models provides a means to create labels for model training. This article provides a short history and review of the annotation and labeling process of medical images, from the traditional manual methods to the newest semiautomated methods that provide a more scalable solution for creating useful models more efficiently. Keywords: Feature Detection, Diagnosis, Semi-supervised Learning © RSNA, 2025.

This study aims to identify distinct imaging subtypes of glioblastoma using multi-modal MRIs from the ReSPOND consortium, providing insights into tumor heterogeneity to inform personalized treatment approaches. We analyzed 3,145 subjects with multi-modal MRIs (T1, T2, T1Gd, FLAIR) from 16 geographically distinct institutions across North America, Europe, and Asia. Radiomic features were extracted using Masked Auto-Encoder (MAE), a deep learning architecture trained through self-supervised learning on 23,608 MRIs across 11-dimensional MRI-derived imaging measures (conventional, diffusion, and perfusion protocols). For clustering analysis, we used features extracted from the four structural modalities available in all subjects. Following feature extraction, we applied ComBat harmonization to remove institutional batch effects, performed dimensionality reduction using cross-validated PCA, and employed K-medoids clustering with consensus analysis (100 random data partitions with a 90/10 split for clustering/validation) to identify stable subtypes, with optimal cluster number determined using adjusted rand index. Additionally, we extracted morphological, intensity, and textural features to characterize the identified subtypes in an interpretable manner. Our analysis revealed three reproducible glioblastoma subtypes. Statistical analysis demonstrated significant differences between subtypes in tumor morphology and spatial location features. Subtype 1 showed centrally located, spherical tumors; Subtype 2 exhibited peripheral, irregular morphologies; Subtype 3 had mixed features (p < 0.001). Key features included right hemisphere ratio, centroid Z position, sphericity, and shape convexity ratio (p < 0.001). Survival analysis indicated median overall survival of 12.3, 14.5, and 16.8 months for Subtypes 1, 2, and 3, respectively (p = 0.042, borderline). Preliminary genomic analysis showed frequent TP53 and RB1 co-mutations, with distinct molecular patterns per subtype. This study demonstrates that multi-modal MRIs can successfully identify glioblastoma subtypes characterized by deep learning derived radiomic features. Future work will focus on validation with comprehensive molecular and tissue-based profiles, survival outcomes, and treatment responses to support personalized medicine.

This report addresses the technical aspects of de-identification of medical images of human subjects and biospecimens, such that re-identification risk of ethical, moral, and legal concern is sufficiently reduced to allow unrestricted public sharing for any purpose, regardless of the jurisdiction of the source and distribution sites. All medical images, regardless of the mode of acquisition, are considered, though the primary emphasis is on those with accompanying data elements, especially those encoded in formats in which the data elements are embedded, particularly Digital Imaging and Communications in Medicine (DICOM). These images include image-like objects such as Segmentations, Parametric Maps, and Radiotherapy (RT) Dose objects. The scope also includes related non-image objects, such as RT Structure Sets, Plans and Dose Volume Histograms, Structured Reports, and Presentation States. Only de-identification of publicly released data is considered, and alternative approaches to privacy preservation, such as federated learning for artificial intelligence (AI) model development, are out of scope, as are issues of privacy leakage from AI model sharing. Only technical issues of public sharing are addressed.

Medical imaging is undergoing a transformation driven by the advent of new, highly effective, machine learning techniques paired with increases in computational capabilities (Cheng et al. 2021; Gilson et al. 2023; Almeida et al. 2024; Krishna et al. 2024). These advanced algorithms have the potential to improve disease detection, diagnosis, prognosis, and treatment outcomes. However, the complexity of machine learning models, the large amounts of curated and annotated data required by some methods, and the potential for bias and error make it challenging for individuals to safely and effectively leverage these methods (Lin et al. 2024; Guo et al. 2024; Xu et al. 2024; Linguraru et al. 2024; Wood et al. 2019). To address these challenges, the American Association of Physicists in Medicine (AAPM), American College of Radiology (ACR), Radiological Society of North America (RSNA), and Society for Imaging Informatics in Medicine (SIIM) have worked together to develop a syllabus detailing a recommended set of competencies for medical imaging professionals interacting with these systems. This guide is aimed at four different personas: users of AI systems, purchasers of AI systems, individuals who provide clinical expertise during the development of AI systems ("clinical collaborators"), and developers of AI systems.1 This is a syllabus, not a curriculum, and is intentional in this scope. Recognizing that individuals may benefit from different presentations of the same material, this work enumerates a series of relevant competencies but does not prescribe, nor offer, a method of instruction (Schuur, Rezazade Mehrizi, and Ranschaert 2021; Garin et al. 2023). By addressing the task-specific demands of each role, this guide will enable medical imaging professionals to utilize machine learning systems more safely and effectively, ultimately improving patient care and outcomes.

Spinal epidural angiolipoma is a very rare neoplasm of the midthoracic spine, comprised of mature adipocytes mixed with abnormal vessels. We report a case of a thoracic dorsal epidural angiolipoma in a 30-year-old pregnant woman (G1P0 at 39 weeks), with back pain and progressive worsening weakness for 2−3 months, who was diagnosed with a T3-8 epidural angiolipoma. The mass was resected after a cesarean delivery. Postoperatively, there was gradual interval improvement of the lower extremity weakness and back pain.

Generative large language models that were fine-tuned on synthetic and MIMIC chest radiograph radiology reports greatly enhanced error detection in radiology reports, demonstrating their potential to serve as powerful tools for medical proofreading.

Purpose The Medical Imaging and Data Resource Center (MIDRC) mRALE Mastermind Grand Challenge fostered the development of artificial intelligence (AI) techniques for the automated assignment of mRALE (modified radiographic assessment of lung edema) scores to portable chest radiographs from patients known to have COVID-19.Approach The challenge utilized 2079 training cases obtained from the publicly available MIDRC data commons, with validation and test cases sampled from not-yet-public MIDRC cases that were inaccessible to challenge participants. The reference standard mRALE scores for the challenge cases were established by a pool of 22 radiologist annotators. Using the MedICI challenge platform, participants submitted their trained algorithms encapsulated in Docker containers. Algorithms were evaluated by the challenge organizers on 814 test cases through two performance assessment metrics: quadratic-weighted kappa and prediction probability concordance.Results Nine AI algorithms were submitted to the challenge for assessment against the test set cases. The algorithm that demonstrated the highest agreement with the reference standard had a quadratic-weighted kappa of 0.885 and a prediction probability concordance of 0.875. Substantial variability in mRALE scores assigned by the annotators and output by the AI algorithms was observed.Conclusions The MIDRC mRALE Mastermind Grand Challenge revealed the potential of AI to assess COVID-19 severity from portable CXRs, demonstrating promising performance against the reference standard. The observed variability in mRALE scores highlights the challenges in standardizing severity assessment. These findings contribute to ongoing efforts to develop AI technologies for potential use in clinical practice and offer insights for the enhancement of COVID-19 severity assessment.

In this retrospective study, a dataset was constructed with two parts. The first part included 1,656 synthetic chest radiology reports generated by GPT-4 using specified prompts, with 828 being error-free synthetic reports and 828 containing errors. The second part included 614 reports: 307 error-free reports between 2011 and 2016 from the MIMIC-CXR database and 307 corresponding synthetic reports with errors generated by GPT-4 on the basis of these MIMIC-CXR reports and specified prompts. All errors were categorized into four types: negation, left/right, interval change, and transcription errors. Then, several models, including Llama-3, GPT-4, and BiomedBERT, were refined using zero-shot prompting, few-shot prompting, or fine-tuning strategies. Finally, the performance of these models was evaluated using the F1 score, 95% confidence interval (CI) and paired-sample t-tests on our constructed dataset, with the prediction results further assessed by radiologists. Using zero-shot prompting, the fine-tuned Llama-3-70B-Instruct model achieved the best performance with the following F1 scores: 0.769 for negation errors, 0.772 for left/right errors, 0.750 for interval change errors, 0.828 for transcription errors, and 0.780 overall. In the real-world evaluation phase, two radiologists reviewed 200 randomly selected reports output by the model. Of these, 99 were confirmed to contain errors detected by the models by both radiologists, and 163 were confirmed to contain model-detected errors by at least one radiologist. Generative LLMs, fine-tuned on synthetic and MIMIC-CXR radiology reports, greatly enhanced error detection in radiology reports.

BACKGROUND:Glioblastoma (GBM) is the most aggressive adult primary brain cancer, characterized by significant heterogeneity, posing challenges for patient management, treatment planning, and clinical trial stratification. METHODS:We developed a highly reproducible, personalized prognostication, and clinical subgrouping system using machine learning (ML) on routine clinical data, magnetic resonance imaging (MRI), and molecular measures from 2838 demographically diverse patients across 22 institutions and 3 continents. Patients were stratified into favorable, intermediate, and poor prognostic subgroups (I, II, and III) using Kaplan-Meier analysis (Cox proportional model and hazard ratios [HR]). RESULTS:The ML model stratified patients into distinct prognostic subgroups with HRs between subgroups I-II and I-III of 1.62 (95% CI: 1.43-1.84, P < .001) and 3.48 (95% CI: 2.94-4.11, P < .001), respectively. Analysis of imaging features revealed several tumor properties contributing unique prognostic value, supporting the feasibility of a generalizable prognostic classification system in a diverse cohort. CONCLUSIONS:Our ML model demonstrates extensive reproducibility and online accessibility, utilizing routine imaging data rather than complex imaging protocols. This platform offers a unique approach to personalized patient management and clinical trial stratification in GBM.

PURPOSE:This study aims to assess whether water exchange rate (kw), a surrogate for blood-brain barrier (BBB) permeability, is associated with functional outcomes in patients with acute ischemic stroke (AIS). METHODS:We studied 22 AIS patients enrolled from 1/2022 to 4/2024 who underwent multi-modal non-contrast imaging on a 3.0-Tesla scanner, including DP-pCASL, DTI, NODDI and MAP imaging. For each parametric map, the intensity and standard deviation (SD) were calculated for the infarcted region. The diffusion maps included were b0, Fractional Anisotropy (FA), Mean Diffusivity (MD), Intra-cellular Volume Fraction (ICVF), Free Water Fraction (FWF), and Orientation Dispersion Index (ODI), q-Space Mean Square Displacement (QMSD), Return-to-Axis Probability (RTAP), Return-to-Plane Probability (RTPP), Return-to-Origin Probability (RTOP), Propagator Anisotropy (PA), and non-Gaussianity (NG). The perfusion-based maps were cerebral blood flow (CBF), arterial transit time (ATT), and kw. The outcome variable was modified Rankin Scale (mRS). RESULTS:Twenty-two patients were included. The average age was 69.5 ± 13.5, the mean NIHSS of 12.4 ± 7.7, and the median infarct of 25.7 (8.4-98.8) ml. Multivariable linear regression identified lower kw (β = -0.029, p = 0.041), longer time to MRI (β = 0.012, p = 0.013) and larger stroke volume (β = 0.014, p = 0.006) as predictive of higher mRS. Higher CBF (β = 0.660, p = 0.003) and RTAP (β = 1.528, p = 0.010), and lower SD RTAP (β = -0.709, p = 0.016), RTPP (β = -2.132, p = 0.006), and NG (β = -1.036, p = 0.011) were identified as most predictive of kw through multivariable linear regression analysis. CONCLUSIONS:Lower kw is predictive of higher mRS in patients with AIS. Increasing CBF and RTAP and lower SD RTAP, RTPP, and NG were correlated with higher kw.

Background and purpose: While mechanical thrombectomy (MT) achieves restoration of cerebral blood flow to the area at risk in patients with acute ischemic stroke (AIS), the influx of blood flow may exacerbate the blood-brain barrier (BBB) disruption and extravasation across the BBB, and it therefore remains unclear how reperfusion impacts the blood-brain barrier integrity. In this study, we use diffusion-prepared pseudocontinuous ASL (DPpCASL) and Neurite Orientation Dispersion and Density Imaging (NODDI) sequence to measure the water exchange rate (kw) in patients who underwent either MT or medical management and determine its impact on the brain tissue microstructure in order to elucidate the impact of MT on BBB complex integrity. Materials and methods: We prospectively enrolled 21 patients with AIS treated at our institution from 10/2021 to 6/2023 who underwent MR imaging at a 3.0-Tesla scanner. Patients underwent DP-pCASl and NODDI imaging in addition to the standard stroke protocol which generated cerebral blood flow (CBF), arterial transit time (ATT), water exchange rate (kw), orientation dispersion index (ODI), intracellular volume fraction (ICVF), and free water fraction (FWF) parametric maps. Results: Of the 21 patients, 11 underwent MT and 10 were treated non-operatively. The average age and NIHSS for the MT cohort and non-MT cohorts were 69.3 +/- 16.6 years old and 15.0 (12.0-20.0), and 70.2 +/- 10.7 (p = 0.882) and 6.0 (3.8-9.0, p = 0.003) respectively. The average CBF, ATT, and kw in the infarcted territory of the MT cohort were 38.2 (18.4-59.6), 1347.6 (1182.5-1842.3), and 107.8 (79.2-140.1) respectively. The average CBF, ATT, and kw in the stroke ROI were 16.0 (8.8-36.6, p = 0.036), 1090.8 (937.1-1258.9, p = 0.013), 89.7 (68.0-122.7, p = 0.314) respectively. Linear regression analysis showed increasing CBF (p = 0.008) and undergoing mechanical thrombectomy (p = 0.048) were significant predictors of increased kw. Conclusion: Using our multimodal non-contrast MRI protocol, we demonstrate that increased CBF and mechanical thrombectomy increased kw, suggesting a better functioning BBB complex. Higher kw suggests less disruption of the BBB complex in the MT cohort.