Normative Baseline for Radiomics in Brain MRI: Evaluating the Robustness, Regional Variations, and Reproducibility on FLAIR Images.

BACKGROUND: Radiomics in neuroimaging has gained momentum as a noninvasive prediction tool not only to differentiate between types of brain tumors, but also to create phenotypic signatures in neurological and neuropsychiatric disorders. However, there is currently little understating about the robustness and reproducibility of radiomic features in a baseline normative population. PURPOSE: To investigate the intra- and interscanner reproducibility, spatial robustness, and sensitivity of radiomics on fluid attenuation inversion recovery (FLAIR) images, which are widely used in neuro-oncology investigations. STUDY TYPE: Retrospective. POPULATION: Three separate datasets of healthy controls: 1) 87 subjects (age range 12-64 years), 2) intrascanner three timepoints, four subjects, and 3) interscanner, eight subjects at three different sites. FIELD STRENGTH/SEQUENCE: T2 -weighted FLAIR at 1.5T and 3.0T. ASSESSMENT: Spatial variance across lobes, and their relation with age/gender, intra- and inter-scanner reproducibility (with and without site harmonization) of radiomics. STATISTICAL TESTS: Analysis of variance (ANOVA), interclass correlation (ICC), coefficient of variation (CoV), Bland-Altman analysis. RESULTS: Analysis of data revealed no differences between genders; however, multiple radiomic features were highly associated with age (P < 0.05). Spatial variability was also evaluated where only 29.04% gray matter and 38.7% white matter features demonstrated an ICC >0.5. Furthermore, the results demonstrated intra-scanner reliability (ICC >0.5); however, inter-scanner reproducibility was poor, with ICC < 0.5 for 82% gray matter and 78.5% white matter features. The inter-scanner reliability improved (ICC < 0.5 for 39.67% gray matter and 38% white matter features) using site-harmonization techniques. DATA CONCLUSION: These findings suggest that, accounting for age, spatial locations in radiomics-based analysis and use of intersite radiomics harmonization is crucial before interpreting these features for pathological inference. Level of Evidence 3. Technical Efficacy Stage 1. J. MAGN. RESON. IMAGING 2021;53:394-407.

Atrophy of cerebellar peduncles in essential tremor: a machine learning-based volumetric analysis.

BACKGROUND: Subtle cerebellar signs are frequently observed in essential tremor (ET) and may be associated with cerebellar dysfunction. This study aims to evaluate the macrostructural integrity of the superior, middle, and inferior cerebellar peduncles (SCP, MCP, ICP) and cerebellar gray and white matter (GM, WM) volumes in patients with ET, and compare these volumes between patients with and without cerebellar signs (ETc and ETnc). METHODS: Forty patients with ET and 37 age- and gender-matched healthy controls were recruited. Atlas-based region-of-interest analysis of the SCP, MCP, and ICP and automated analysis of cerebellar GM and WM volumes were performed. Peduncular volumes were employed in a multi-variate classification framework to attempt discrimination of ET from controls. RESULTS: Significant atrophy of bilateral MCP and ICP and bilateral cerebellar GM was observed in ET. Cerebellar signs were present in 20% of subjects with ET. Comparison of peduncular and cerebellar volumes between ETnc and ETc revealed atrophy of right SCP, bilateral MCP and ICP, and left cerebellar WM in ETc. The multi-variate classifier could discriminate between ET and controls with a test accuracy of 86.66%. CONCLUSIONS: Patients with ET have significant atrophy of cerebellar peduncles, particularly the MCP and ICP. Additional atrophy of the SCP is observed in the ETc group. These abnormalities may contribute to the pathogenesis of cerebellar signs in ET. KEY POINTS: • Patients with ET have significant atrophy of bilateral middle and inferior cerebellar peduncles and cerebellar gray matter in comparison with healthy controls. • Patients of ET with cerebellar signs have significant atrophy of right superior cerebellar peduncle, bilateral middle and inferior cerebellar peduncle, and left cerebellar white matter in comparison with ET without cerebellar signs. • A multi-variate classifier employing peduncular volumes could discriminate between ET and controls with a test accuracy of 86.66%.

Federated learning enables big data for rare cancer boundary detection.

Although machine learning (ML) has shown promise across disciplines, out-of-sample generalizability is concerning. This is currently addressed by sharing multi-site data, but such centralization is challenging/infeasible to scale due to various limitations. Federated ML (FL) provides an alternative paradigm for accurate and generalizable ML, by only sharing numerical model updates. Here we present the largest FL study to-date, involving data from 71 sites across 6 continents, to generate an automatic tumor boundary detector for the rare disease of glioblastoma, reporting the largest such dataset in the literature (n = 6, 314). We demonstrate a 33% delineation improvement for the surgically targetable tumor, and 23% for the complete tumor extent, over a publicly trained model. We anticipate our study to: 1) enable more healthcare studies informed by large diverse data, ensuring meaningful results for rare diseases and underrepresented populations, 2) facilitate further analyses for glioblastoma by releasing our consensus model, and 3) demonstrate the FL effectiveness at such scale and task-complexity as a paradigm shift for multi-site collaborations, alleviating the need for data-sharing.