PURPOSE:The purpose of this study was to prospectively evaluate the contribution of intravoxel incoherent motion (IVIM) and diffusion kurtosis imaging (DKI) for predicting outcome after ischemic stroke. MATERIALS AND METHODS:Patients with acute ischemic stroke who underwent brain MRI at 3 Tesla, including a multi-b diffusion-weighted imaging sequence were prospectively included. Mean, maximum (max) and minimum (min) of diffusion parameters, including apparent diffusion coefficient (ADC), D, D*, f, f·D* and K were extracted from the infarct core and normal-appearing parenchyma. The primary endpoint was 3-month modified Rankin scale (mRS). Associations were tested by univariable ordinal logistic regression with false discovery rate correction. Model performance was assessed using area under the receiver operating characteristic curve (AUC) analysis and the ranked probability score. RESULTS:A total of 157 patients were included. There were 83 men and 74 women with a mean age of 70 ± 16 (standard deviation) years (range: 23-100 years). In univariable analyses, ADCmin in the infarct core was the strongest predictor of worse functional outcome at 3 months (OR, 0.37; P < 0.001). Within the infarct core, nine other parameters were significantly associated with mRS (one ADC, six IVIM and two DKI metrics). In the normal-appearing parenchyma, eight parameters (two ADC, four IVIM and two DKI metrics) were significantly associated with outcome. In multivariable analysis, the model combining ADCmin and Kmin in the core with ADCmax in the normal-appearing parenchyma provided the highest prognostic performance, with an AUC of 0.81 (95 % confidence interval [CI]: 0.77-0.86) compared to ADCmin alone (AUC, 0.75; 95 % CI: 0.67-0.82) (P < 0.0001). CONCLUSION:A combination of advanced diffusion MRI parameters could predict functional outcome in patients with ischemic stroke with higher precision than standard ADC alone.

To compare conventional speech recognition (CSR) and a general-purpose large language model (LLM) in radiology reports, focusing on generation times and errors. In this prospective, multicenter study, five radiologists produced 200 reports using CSR and 200 using a general-purpose LLM with in-built speech recognition during routine clinical practice. Generation times were recorded. Errors were evaluated qualitatively and quantitatively using Levenshtein distance. Mann–Whitney U-test was used to compare quantitative variables, and chi-square test for categorical variables. No patient-identifying or clinical information was uploaded to the LLM. 301/400 (75.3

PURPOSE:The purpose of this study was to develop a machine learning-based algorithm based on a combination of magnetic resonance imaging (MRI) and color-Doppler ultrasound (CDUS) to characterize lacrimal gland lesions. MATERIALS AND METHODS:All patients with a lacrimal gland lesion who underwent MRI examination and CDUS between 2014 and 2025 were retrospectively included. Thirty-four imaging features were systematically assessed. A machine learning algorithm was trained with repeated nested cross-validation (RNCV) using random forest classifiers. Shapley additive explanations values were used to assess feature contributions. Simplified models using top 5 and top 10 best features were also developed. Diagnostic performance of the models was assessed using area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (PR AUC), balanced accuracy, precision, sensitivity, specificity, Brier score, Matthew's correlation coefficient and F1-score. RESULTS:One hundred patients (mean age, 49.6 years ± 17.8 [standard deviation] years) with 130 lesions (101 non-epithelial (NEL) and 29 epithelial (EL); 45 malignant) were included. The random forest binary machine learning model yielded 75.9% sensitivity (95% confidence interval [CI]: 39-100), 86.0% specificity (95% CI: 62.4-100), and an AUC of 0.883 (95% CI: 0.692-1.0) for differentiating between malignant and benign lesions and 73.2% sensitivity (95% CI: 33.5-100), 92.9% specificity (95% CI: 69.9-100), and an AUC of 0.93 (95% CI: 0.683-1) for differentiating between EL and NEL. In multiclass analysis (benign NEL, benign EL, malignant NEL and malignant EL), the random forest yielded a macro-averaged AUC of 0.857 (95% CI: 0.722-0.972) for the all-features model. A 5-top features signature comprising apparent diffusion coefficient and resistance index values, echogenicity, age and lesion type (infiltrative vs. well-delineated mass), yielded an AUC of 0.785 (95% CI: 0.641-0.941) to distinguish between the four classes. CONCLUSION:A combination of MRI and CDUS features demonstrated high diagnostic performance for characterizing lacrimal gland lesions. A simplified 5-feature signature showed similar diagnostic performance compared to the all-features model and warrants prospective multicenter validation for clinical application.

OBJECTIVES:Synthetic magnetic resonance imaging (MRI) is a quantitative imaging technique that has shown promise in brain imaging but has not yet been evaluated for assessing the optic nerves. Our study aimed to investigate its diagnostic performance in this context. MATERIALS AND METHODS:We retrospectively evaluated synthetic MRI's performance in detecting optic nerve hypersignals in 65 patients who underwent synthetic MRI covering the optic nerves from March 2023 to February 2025 in a single tertiary center. Diagnostic performance for optic nerve hypersignals was assessed using conventional T2 and/or FLAIR-weighted images with fat saturation as the reference standard. Quantitative T2 and proton density (PD) values were compared between optic nerves exhibiting hypersignals on synthetic MRI and those without any hypersignals. The detection rate of optic nerve hypersignals in patients with a diagnosis of acute optic neuritis was evaluated using synthetic MRI, both overall and for each individual synthetic contrast. For the qualitative analysis, sensitivity, specificity, and accuracy were each calculated with a 95% CI using the exact binomial (Clopper-Pearson) method. Quantitative differences in T2 and PD values were assessed using the Cohen d to evaluate effect size, and statistical significance was determined by the Wilcoxon rank-sum test. RESULTS:Synthetic MRI showed good overall diagnostic performance for optic nerve hypersignals, with sensitivity, specificity, and accuracy of 71.4% [0.513-0.868], 97.1% [0.916-0.994], and 91.5% [0.854-0.957], respectively. Quantitative analysis revealed significantly higher median T2 (66.29 vs. 72.4 ms) and proton density (72.22 vs. 86.51) values in optic nerves exhibiting hypersignals compared with those without ( P <0.001 for both). For acute optic neuritis specifically, 6 out of 7 (85.7%) were correctly identified in synthetic MRI. Confidence scores did not significantly differ between patients with optic nerve hypersignals and those without. CONCLUSIONS:Synthetic MRI showed promising results in detecting abnormal signals in the optic nerves, suggesting its potential role in their clinical evaluation.

BACKGROUND:Non-small cell lung cancer (NSCLC) staging relies on accurate assessment of mediastinal lymph nodes. This study investigates the utility of radiomic features derived from contrast-enhanced thoracic CT scans in predicting malignancy in clinically positive (cN+) mediastinal lymph nodes. METHODS:A retrospective cohort of 110 NSCLC patients with cN + who underwent surgical resection was analyzed. 3D segmentations of up to three lymph nodes per patient were performed. Radiomic features, encompassing heterogeneity measures, were extracted. A radiomics model was constructed using a random forest and XGboost algorithm. To ensure robustness and minimize bias, 100 iterations of data splitting were conducted to create distinct training and test sets for reproducibility and statistical reliability. RESULTS:The radiomics model achieved an area under the curve (AUC) of 0.703 for Forest model and 063 for XGboost model. Three recurrent features in the radiomic signatures, "RootMeanSquared", "Grey Level Co-Occurrence Matrix Imc2", and "Grey Level Run Length Matrix RunEntropy", highlighted the importance of nodal heterogeneity features in the model. CONCLUSION:Radiomic features extracted from contrast-enhanced thoracic CT scans could predict malignancy in cN+ mediastinal lymph nodes of NSCLC patients (AUC = 0.703). Three features of heterogeneity were recurrent in radiomic signature, emphasizing the potential importance of incorporating nodal heterogeneity criteria in addition to size assessment.

Objectifs L’IRM synthétique est une technique quantitative prometteuse en imagerie cérébrale [1-4], mais qui n’a jusqu’à présent pas été évaluée pour l’exploration des nerfs optiques. Cette étude a pour objectif d’en analyser les performances diagnostiques dans ce contexte. Matériels et méthodes Nous avons conduit une étude rétrospective incluant 65 patients ayant bénéficié d’une IRM synthétique couvrant les nerfs optiques entre mars 2023 et février 2025 dans un centre tertiaire. Le standard de référence reposait sur les séquences T2 et/ou FLAIR conventionnelles. Les valeurs quantitatives de T2 et de densité de protons (DP) ont été comparées entre nerfs optiques avec et sans hypersignal. La détection des hypersignaux liés aux névrites optiques aiguës a été spécifiquement analysée. La sensibilité, la spécificité et la précision diagnostique ont été calculées avec un intervalle de confiance à 95 % (méthode exacte de Clopper &Pearson). Les différences quantitatives ont été évaluées par le test de Wilcoxon et la taille d’effet estimée par le d de Cohen. Résultats L?IRM synthétique a montré une bonne performance diagnostique pour la détection des hypersignaux du nerf optique (sensibilité 71.4 % [0.513?0.868], spécificité 97.1 % [0.916?0.994], précision 91.5 % [0.854?0.957]). Les valeurs médianes de T2 (72.4 vs 66.3 ms) et de DP (86.5 vs 72.2) étaient significativement plus élevées dans les nerfs présentant un hypersignal (p < 0.001). Parmi les 7 cas de névrite optique aiguë, 6 (85.7 %) ont été correctement détectés. Les scores de confiance ne différaient pas entre nerfs avec ou sans hypersignal. Conclusion L’IRM synthétique est une approche prometteuse pour la détection des anomalies de signal du nerf optique et pourrait avoir un rôle dans leur évaluation clinique.

OBJECTIVE:This study evaluated whether radiomics analysis of standard hand radiographs could provide an automated and objective assessment of structural severity in hand osteoarthritis (HOA) compared to Kellgren-Lawrence (KL) scores, ranging from 0 (normal) to 4 (severe). DESIGN:We conducted a retrospective study using baseline data from the DIGICOD (DIGital Cohort Osteoarthritis Design) cohort, including patients with HOA. Standard posteroanterior radiographs were segmented semi-automatically using a U-Net model with manual correction. Radiomics features describing intensity, shape, and texture were extracted for each joint and reduced after filtering and correlation suppression. The cohort was split into training (80%) and test sets (20%) using stratified sampling. Random Forest classifiers were trained to detect structural involvement (KL ≥2), severe disease (KL 3-4), and predict multiclass KL grades (0-1, 2, 3, 4) at the joint level. RESULTS:379 radiographs were analyzed. Detection of structural involvement (KL ≥2) achieved an AUC of 0.81 [95% CI: 0.79-0.83], with high sensitivity (89%) but low specificity (59%). Severe disease detection (KL 3-4) reached an AUC of 0.83 [95% CI: 0.81-0.85], with very good sensitivity (80%) and good specificity (70%). Multiclass KL prediction showed lower performance (macro-averaged AUC 0.76; accuracy 56%), reflecting challenges in distinguishing intermediate (KL 2-3) and severe (KL 4) grades. CONCLUSIONS:This is the first study applying radiomics to standard hand radiographs for automated and objective scoring of radiographic severity in HOA. The model showed good performance detecting structural damage, supporting radiomics as a potential tool to reduce reliance on subjective visual grading.

Over a third of minor stroke patients experience post-stroke cognitive impairment (PSCI), but no validated tools exist to identify at-risk patients early. This study investigated whether disconnection features derived from infarcts and white matter hyperintensities (WMH) could serve as markers for short- and long-term cognitive decline in first-ever minor ischemic stroke patients. First-ever minor ischemic stroke patients (NIHSS ≤ 7) were prospectively followed at 72-h, 6 months, and 36 months post-stroke with cognitive tests and brain MRI. Infarct and WMH volumes were semi-automatically assessed on DWI and FLAIR sequences. Bayesian tract-based disconnection models estimated remote pathological effects of infarcts and WMH. Associations between disconnection features and cognitive outcomes were analyzed using canonical correlation analyses, adjusted for age, education, and multiple comparisons. Among 105 patients (31% female, mean age 63 ± 12 years), infarct volume averaged 10.28 ± 17.10 cm 3 and predominantly involved the middle cerebral artery territory (83%). WMH burden was higher in frontal periventricular white matter. Infarct-based features did not significantly relate to PCSI. However, a WMH-derived disconnection factor, involving commissural and frontal tracts, and the right superior longitudinal fasciculus, was significantly associated with PSCI at 6 months (OR = 9.96, p value = 0.02) and 36 months (OR = 12.27, p value = 0.006), particularly in executive/attention, language, and visuospatial domains. This factor, unrelated to WMH volume, outperformed demographic and clinical predictors of PSCI. WMH-induced disconnection may be associated with short- and long-term PSCI in minor stroke. Routine MR-derived features could identify at-risk patients for rehabilitation trials.

PURPOSE:The purpose of this study was to assess the benefit of a deep learning-based image reconstruction (DLBIR) for improving image quality in orbital magnetic resonance imaging (MRI) at 3 Tesla (T). MATERIALS AND METHODS:Seventy-one patients (48 women and 23 men) with a mean age of 52 ± 19.5 (standard deviation [SD]) years (age range: 7-90 years) who underwent MRI examination of the orbit at 3 T between January and June of 2024, were included in the study. Coronal T2-weighted MR images obtained in 70 patients and post-contrast fat-saturated (FS) coronal T1-weighted MR images obtained in 25 patients, were reconstructed with and without DLBIR, resulting in four imaging sets. Two radiologists independently and blindly measured the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of the optic nerves on the four imaging sets. Image quality and orbital abnormalities were assessed using a standardized 5-point Likert scale. Comparisons between MR images obtained with and without DLBIR were performed using Wilcoxon test for ordinal and quantitative variables and McNemar test for paired binary data. RESULTS:SNR and CNR of coronal T2-weighted MR images were significantly greater using DLBIR (26.67 ± 9.03 [SD], and 14.87 ± 10.31 [SD], respectively) than without DLBIR (18.91 ± 7.28 [SD], and 9.78 ± 8.47 [SD], respectively) (P < 0.001). There were no differences in SNR and CNR between post-contrast FS T1-weighted images obtained with DLBIR (85.56 ± 63.13 [SD], and 64 ± 41.38 [SD], respectively) and those obtained without DLBIR (91.36 ± 48.49 [SD], and 43.25 ± 20.4 [SD], respectively) (P = 0.35, and P = 0.14, respectively). Qualitatively, good-to-excellent image quality was obtained more frequently with DLBIR than without DLBIR for T2-weighted and post-contrast FS T1-weighted images with respect to optic nerve sharpness (67 % vs. 16 %, and 8 % vs. 0 %, respectively), brain sharpness (90 % vs. 6 %, and 68 % vs. 4 %, respectively), and overall image quality (73 % vs. 1 % and 36 % vs. 0 %, respectively) (all P ≤ 0.001). No significant differences in the detection rates of orbital abnormalities were found between MR images obtained with and without DLBIR, including optic nerve hyperintensity (34 % vs. 31 %, respectively; P = 0.16) and optic nerve atrophy (33 % for both) on T2-weighted images, and optic nerve enhancement on post-contrast FS T1-weighted images (16 % for both). CONCLUSION:DLBIR significantly improves image quality of MRI examinations of the orbit at 3 T, without losing clinically relevant information.

Synthetic magnetic resonance imaging (MRI) is a quantitative imaging technique that has shown promise in brain imaging but has not yet been evaluated for assessing the optic nerves. Our study aimed to investigate its diagnostic performance in this context. We retrospectively evaluated synthetic MRI’s performance in detecting optic nerve hypersignals in 65 patients who underwent synthetic MRI covering the optic nerves from March 2023 to February 2025 in a single tertiary center. Diagnostic performance for optic nerve hypersignals was assessed using conventional T2 and/or FLAIR-weighted images with fat saturation as the reference standard. Quantitative T2 and proton density (PD) values were compared between optic nerves exhibiting hypersignals on synthetic MRI and those without any hypersignals. The detection rate of optic nerve hypersignals in patients with a diagnosis of acute optic neuritis was evaluated using synthetic MRI, both overall and for each individual synthetic contrast. For the qualitative analysis, sensitivity, specificity, and accuracy were each calculated with a 95% CI using the exact binomial (Clopper-Pearson) method. Quantitative differences in T2 and PD values were assessed using the Cohen d to evaluate effect size, and statistical significance was determined by the Wilcoxon rank-sum test. Synthetic MRI showed good overall diagnostic performance for optic nerve hypersignals, with sensitivity, specificity, and accuracy of 71.4% [0.513-0.868], 97.1% [0.916-0.994], and 91.5% [0.854-0.957], respectively. Quantitative analysis revealed significantly higher median T2 (66.29 vs. 72.4 ms) and proton density (72.22 vs. 86.51) values in optic nerves exhibiting hypersignals compared with those without (P<0.001 for both). For acute optic neuritis specifically, 6 out of 7 (85.7%) were correctly identified in synthetic MRI. Confidence scores did not significantly differ between patients with optic nerve hypersignals and those without. Synthetic MRI showed promising results in detecting abnormal signals in the optic nerves, suggesting its potential role in their clinical evaluation.

PURPOSE:The purpose of this study was to develop a computed tomography (CT)-based radiomic model and evaluate its performance in discriminating between different infectious agents in immunocompromised patients with pulmonary infection. MATERIALS AND METHODS:This single-center retrospective study included immunocompromised patients with pulmonary infections presenting as focal, nodular lung lesion(s) on CT from 2012 to 2023. Thirteen clinical and CT semantic features were collected. Three-dimensional segmentation of the main lung lesion was performed on CT images, followed by radiomic feature extraction. The dataset was divided into training (80 %) and test (20 %) sets. Radiomic, clinical/semantic, and combined models were built using a multi-class random forest classifier and a 5-fold stratified cross-validation in the training set and tested in the test set. The performance of each model was evaluated using the area under the receiver operating characteristic curve (AUC). RESULTS:One hundred and ninety-six patients with aspergillosis (n = 123), tuberculosis (n = 41), or "mucormycosis/nocardiosis" (n = 32) were included. There were 131 men and 65 women with a median age of 67 years (age range: 20-95 years). The mean AUC of the radiomic model was 0.84 (95 % confidence interval [CI]: 0.66, 0.98), whereas the mean AUC of the clinical/semantic model was 0.66 [95 % CI: 0.46, 0.84]). The mean AUC of the combined model was similar to that of the radiomic model (0.82 [95 % CI: 0.63, 0.97]). The AUCs of the radiomic model (0.89 [95 % CI: 0.74, 1.00]) and of the combined models (0.93 [95 % CI: 0.83, 0.99]) for discriminating tuberculosis from the other two classes were significantly higher than the AUC of the clinical/semantic model (0.62 [95 % CI: 0.41, 0.81]) (P = 0.046 and 0.003, respectively). CONCLUSIONS:In our data set, the radiomic model performs better than a clinical/semantic model in differentiating tuberculosis from other nodular lung lesions. These encouraging results highlight the potential role of quantitative analysis in contributing to the diagnosis of pulmonary infections in immunocompromised patients.

BACKGROUND:Giant cell arteritis (GCA) is the leading vasculitis threatening vision in adults aged ≥ 50 years; permanent vision loss may occur within the first few days after symptom onset. We assessed the impact of a fast-track pathway (FTP) for early diagnosis and treatment of giant cell arteritis in terms of hospitalization patterns and cost-effectiveness. METHODS:We conducted a retrospective, single-center medico-economic study of consecutive patients referred to a neuro-ophthalmology tertiary center between Nov 1, 2016, and Dec 31, 2022. GCA was defined by ≥ 3 American College of Rheumatology criteria plus a positive temporal-artery biopsy or vascular imaging. An FTP-24/7 access to internal medicine specialists, priority magnetic-resonance imaging, and protocol-driven corticosteroid initiation-was launched on Nov 1, 2018. Demographic, clinical, biological, care-pathway, and cost data were compared before (pre-FTP) and after (post-FTP) implementation. Continuous variables were analyzed with two-sample t tests or Wilcoxon rank-sum tests; categorical variables with χ² or Fisher's exact test. FINDINGS:We included 135 patients (mean age 76 ± 8 years, 61% women): 23 pre-FTP and 112 post-FTP. Baseline characteristics were similar between groups. Compared with the pre-FTP period, the FTP reduced full hospitalizations (62% [69/112] vs 96% [22/23]; p < 0.01) and increased day-hospital or outpatient management (39% vs 4%; p < 0.01). More patients received treatment within one month of symptom onset (54% vs 22%; p < 0.01). Final visual acuity improved (median 2.0 vs 2.6 logMAR; p < 0.01), while cumulative intravenous corticosteroid exposure was significantly reduced (1679 ± 760 mg vs 2295 ± 1055 mg; p = 0.02). Reliance on temporal-artery biopsy fell (17% vs 91%; p < 0.01), owing to a four-fold rise in diagnostic MRI use. Mean total medical costs decreased by €814 per patient (€3672 ± 2861 vs €4486 ± 3193), although this difference did not reach statistical significance (p = 0.23). INTERPRETATION:A dedicated fast-track pathway for suspected GCA enables prompt, largely ambulatory care, halves unnecessary full hospitalizations, speeds treatment initiation, improves visual prognosis, and lowers overall expenditure. These findings support wider adoption of imaging-driven FTPs to mitigate the growing clinical and economic burden of GCA.

Radiology reports, typically recorded as unstructured free text or with varying levels of structuration, contain critical information on tumor evolution but remain difficult to mine for care optimization or research without advanced language processing. We evaluated 15 open-source Large Language Models (LLMs) for classifying tumor evolution from French imaging reports, using a gold-standard corpus of 310 cases. We tested models with varied architecture, hyperparameter configuration and prompting strategy, and compared them with rule-based and BERT-based baselines. We systematically assessed development time and carbon emissions. Properly selected and configured, LLMs outperformed state-of-the-art baselines without requiring large manually annotated datasets, but used substantial computational resources. In contrast, fine-tuned BERT models, trained on high-quality annotations, achieved only slightly lower performance at reduced hardware and computational costs. Our results highlight a trade-off between human annotation effort and computational infrastructure, offering insight for transforming unstructured clinical reports into structured, actionable data.

To develop a deep learning (DL) model for the detection of spinal cord (SC) multiple sclerosis (MS) lesions from both sagittal T2 and short tau inversion recovery (STIR) sequences and to investigate whether such a model could improve the performance of clinicians in detecting SC lesions. A DL tool was developed based on SC sagittal T2 and STIR acquisitions from the imaging database of the French MS registry (OFSEP), including retrospective data from 40 different scanners. A multi-reader study based on retrospective data was performed between December 2023 and June 2024 to compare the performance of 20 clinicians in interpreting upper and lower SC acquisitions with and without the use of the tool. A ground truth was established by three experts. Sensitivity, precision, and inter-reader variability were evaluated. We included 50 patients (39 females, median age: 41 years [range: 15–67]) with SC MRI acquired between February 2017 and December 2022. When reading with the tool, the clinicians’ mean sensitivity to detect SC lesions improved (from 74.3

Les tumeurs orbitaires représentent un défi diagnostique en raison de leurs localisations variées et de leurs nombreuses histopathologies. Au cours des dernières années, les avancées en imagerie ont amélioré le diagnostic mais la classification demeure difficile. L'application de l'intelligence artificielle en radiologie et en ophtalmologie a montré des résultats prometteurs. L'objectif principal de cette étude était de développer et d'évaluer les performances de l'apprentissage automatique dans l'identification des tumeurs orbitaires malignes à partir de l'IRM 3 Tesla multiparamétrique. Cette étude prospective monocentrique a inclus des patients atteints de masses orbitaires explorées par IRM 3 Tesla avant toute chirurgie entre décembre 2015 et mai 2021. Nous avons utilisé un modèle Random Forest avec une validation croisée stratifiée imbriquée, en utilisant différentes combinaisons de variables explicatives. Les valeurs SHAP (SHapley Additive exPlanations) ont été utilisées pour évaluer les contributions des variables. Plusieurs métriques ont évalué les performances du modèle. Nous avons analysé 113 patients (50,4 % de femmes, 49,6 % d'hommes), avec un âge moyen de 51,5 ans [19-88]. Parmi les huit modèles évalués, celui qui incorporait l'ensemble des 46 variables explicatives (morphologie, DWI, DCE et IVIM) obtenait une AUC de 0,9 [0,73-0,99], tandis que le modèle "signature à 10 variables" obtenait une AUC de 0,88 [0,71-0,99]. Les dix variables les plus influentes pour le modèle de Random Forest comprenaient trois variables quantitatives d'IVIM, quatre variables quantitatives DCE, une variable quantitative de DWI, une variable qualitative de DWI qualitative et l'âge [1-5]. Ce travail suggère que l'apprentissage automatique, combinant des données multiparamétriques d'IRM, incluant la DCE, la DWI, l'IVIM et l'imagerie morphologique, peut fournir des modèles performants pour la classification des tumeurs orbitaires. Le modèle "signature à 10 variables" peut être préféré en raison de ses performances solides, de sa simplicité et du principe de parcimonie.