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.

Computational competitions are the standard for benchmarking medical image analysis algorithms, but they typically use small curated test datasets acquired at a few centers, leaving a gap to the reality of diverse multicentric patient data. To this end, the Federated Tumor Segmentation (FeTS) Challenge represents the paradigm for real-world algorithmic performance evaluation. The FeTS challenge is a competition to benchmark (i) federated learning aggregation algorithms and (ii) state-of-the-art segmentation algorithms, across multiple international sites. Weight aggregation and client selection techniques were compared using a multicentric brain tumor dataset in realistic federated learning simulations, yielding benefits for adaptive weight aggregation, and efficiency gains through client sampling. Quantitative performance evaluation of state-of-the-art segmentation algorithms on data distributed internationally across 32 institutions yielded good generalization on average, albeit the worst-case performance revealed data-specific modes of failure. Similar multi-site setups can help validate the real-world utility of healthcare AI algorithms in the future.

The aim was to predict survival of glioblastoma at 8 months after radiotherapy (a period allowing for completing a typical course of adjuvant temozolomide), by applying deep learning to the first brain MRI after radiotherapy completion. Retrospective and prospective data were collected from 206 consecutive glioblastoma, isocitrate dehydrogenase -wildtype patients diagnosed between March 2014 and February 2022 across 11 UK centers. Models were trained on 158 retrospective patients from 3 centers. Holdout test sets were retrospective (n = 19; internal validation), and prospective (n = 29; external validation from 8 distinct centers). Neural network branches for T₂-weighted and contrast-enhanced T₁-weighted inputs were concatenated to predict survival. A nonimaging branch (demographics/MGMT/treatment data) was also combined with the imaging model. We investigated the influence of individual MR sequences; nonimaging features; and weighted dense blocks pretrained for abnormality detection. The imaging model outperformed the nonimaging model in all test sets (area under the receiver-operating characteristic curve, AUC P = .038) and performed similarly to a combined imaging/nonimaging model (P > .05). Imaging, nonimaging, and combined models applied to amalgamated test sets gave AUCs of 0.93, 0.79, and 0.91. Initializing the imaging model with pretrained weights from 10 000s of brain MRIs improved performance considerably (amalgamated test sets without pretraining 0.64; P = .003). A deep learning model using MRI images after radiotherapy reliably and accurately determined survival of glioblastoma. The model serves as a prognostic biomarker identifying patients who will not survive beyond a typical course of adjuvant temozolomide, thereby stratifying patients into those who might require early second-line or clinical trial treatment.

Alcohol-related cerebellar degeneration is one of the commonest acquired forms of cerebellar ataxia. The exact pathogenic mechanisms by which alcohol leads to cerebellar damage remain unknown. Possible autoreactive immune mediated mechanisms have not been explored previously. In this study, we aim to investigate the potential role of alcohol-induced immune mediated cerebellar degeneration. Patients with ataxia and a history of alcohol misuse were recruited from the Ataxia and Hepatology tertiary clinics at Sheffield Teaching Hospitals NHS Trust. We determined the pattern of cerebellar involvement both on clinical (SARA score) and imaging (MRI volumetry and MR spectroscopy) parameters. In addition, HLA genotyping, serological markers for gluten-related disorders and serological reactivity on rat cerebellar tissue using indirect immunohistochemistry were assessed. Thirty-eight patients were included in the study all of whom had ataxia. The gait (97 %), stance (89 %) and heel-shin slide (89 %) were the predominant SARA elements affected. MRI volumetric and spectroscopy techniques demonstrated significant structural, volumetric and functional deficits of the cerebellum with particular involvement of the cerebellar vermis. Circulating anti-gliadin antibodies were detected in 34 % patients vs. 12 % in healthy controls. Antibodies to transglutaminase 6 (TG6) were detected in 39 % of patients and 4 % of healthy control subjects. Using immunohistochemistry, Purkinje cell and/or granular layer reactivity was demonstrated in 71 % of patient sera. Alcohol induced tissue injury to the CNS leading to cerebellar degeneration may also involve immune mediated mechanisms, including sensitisation to gluten.

Magnetic Resonance Imaging (MRI) exhibits scanner dependent contrast, which limits generalisability of radiomics and machine-learning for radiation oncology. Current deep-learning harmonisation requires paired data, retraining for new scanners and often suffers from geometry-shift which alters anatomical information. The aim of this study was to investigate style-blind auto-encoders for MRI harmonisation to accommodate unpaired training data, avoid geometry-shift and harmonise data from previously unseen scanners. A style-blind auto-encoder, using adversarial classification on the latent-space, was designed for MRI harmonisation. The public CC359 T1-w MRI brain dataset includes six scanners (three manufacturers, two field strengths), of which five were used for training. MRI from all six (including one unseen) scanner were harmonised to common contrast. Harmonisation extent was quantified via Kolmogorov-Smirnov testing of residual scanner dependence of 3D radiomic features, and compared to WhiteStripe normalisation. Anatomical content preservation was measured through change in structural similarity index on contrast-cycling (δSSIM). The percentage of radiomics features showing statistically significant scanner-dependence was reduced from 41% (WhiteStripe) to 16% for white matter and from 39% to 27% for grey matter. δSSIM < 0.0025 on harmonisation and de-harmonisation indicated excellent anatomical content preservation. Our method harmonised MRI contrast effectively, preserved critical anatomical details at high fidelity, trained on unpaired data and allowed zero-shot harmonisation. Robust and clinically translatable harmonisation of MRI will enable generalisable radiomic and deep-learning models for a range of applications, including radiation oncology treatment stratification, planning and response monitoring.

Glioblastoma (GBM) has the typical radiological appearance (TRA) of a centrally necrotic, peripherally enhancing tumor with surrounding edema. The objective of this study was to determine whether the developing GBM displays a spectrum of imaging changes detectable on routine clinical imaging prior to TRA GBM. Patients with pre-operative imaging diagnosed with GBM (1 January 2014–31 March 2022) were identified from a neuroscience center. The imaging was reviewed by an experienced neuroradiologist. Imaging patterns preceding TRA GBM were analyzed. A total of 76 out of 555 (14%) patients had imaging preceding TRA GBM, 57 had solitary lesions, and 19 had multiple lesions (total = 84 lesions). Here, 83% of the lesions had cortical or cortical/subcortical locations. The earliest imaging features for 84 lesions were T2 hyperintensity/CT low density (n = 18), CT hyperdensity (n = 51), and T2 iso-intensity (n = 15). Lesions initially showing T2 hyperintensity/CT low density later showed T2 iso-intensity. When CT and MRI were available, all CT hyperdense lesions showed T2 iso-intensity, reduced diffusivity, and the following enhancement patterns: nodular 35%, solid 29%, none 26%, and patchy peripheral 10%. The mean time to develop TRA GBM from T2 hyperintensity was 140 days and from CT hyperdensity was 69 days. This research suggests that the developing GBM shows a spectrum of imaging features, progressing through T2 hyperintensity to CT hyperdensity, T2 iso-intensity, reduced diffusivity, and variable enhancement to TRA GBM. Red flags for non-TRA GBM lesions are cortical/subcortical CT hyperdense/T2 iso-intense/low ADC. Future research correlating this imaging spectrum with pathophysiology may provide insight into GBM growth patterns.