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.

Many tumors including head and neck squamous cell carcinoma (HNSCC) spread along the lymphatic network. Current imaging modalities can only detect sufficiently large metastases. Therefore, adjacent lymph node levels (LNL) are irradiated electively since they may harbor microscopic tumors. We apply Bayesian Networks (BN) to model lymphatic tumor progression. The model can subsequently be used to personalize the risk estimation of microscopic lymph node metastases in newly diagnosed patients based on their distribution of macroscopic metastases. A BN is a graphical representation of a joint probability distribution. We represent LNLs by binary random variables corresponding to the BN nodes. Each LNL is associated with a hidden microscopic state and an observed macroscopic state (e.g. ¹⁸ F-FDG-PET/CT imaging). The primary tumor is represented by network input nodes. We demonstrate the concept for early T-stage oropharyngeal carcinomas and their spread to ipsilateral lymph node levels (LNL) Ib to IV. We show that the BN parameters can be efficiently learnt by merging pathology findings on microscopic tumor progression (which is limited to a few published studies) and imaging data on macroscopic tumor progression such as CT and ¹⁸ F-FDG-PET (which are widely available in clinical practice). The trained network can be used to quantify how the distribution of macroscopic metastases impacts the probability of microscopic involvement of the remaining LNLs. The analysis suggests that the risk of microscopic involvement of level IV exceeds 5% only if level III harbors metastases. Excluding level IV from the elective CTV for other patients would reduce the integral dose delivered to the patient and potentially reduce acute and late side effects.

We provide a dataset on lymph node metastases in 968 patients with newly diagnosed head and neck squamous cell carcinoma (HNSCC). All patients received neck dissection and we report the number of metastatic versus investigated lymph nodes per lymph node level (LNL) for every individual patient. Additionally, clinicopathological factors including T-category, primary tumor subsite (ICD-O-3 code), age, and sex are reported for all patients. The data is provided as three datasets: Dataset 1 contains 373 HNSCC patients treated at Centre Léon Bérard (CLB), France, with primary tumor location in the oral cavity, oropharynx, hypopharynx, and larynx. Dataset 2 contains 332 HNSCC patients treated at the Inselspital, Bern University Hospital (ISB), Switzerland with primary tumor location in the oral cavity, oropharynx, hypopharynx, and larynx. For these patients, additional information is provided including lateralization of the primary tumor, size and location of the largest metastases, and clinical involvement based on computed tomography (CT), magnetic resonance imaging (MRI), and/or 18FDG-positron emission tomography (PET/CT) imaging. Dataset 3 consists of 263 oropharyngeal SCC patients underlying a previous publication by Bauwens et al. [1], which were treated at CLB. For these patients, additional information including HPV status, lateralization of the primary tumor and clinically diagnosed lymph node involvement is provided. The data may be used to quantify the probability of occult lymph node metastases in each LNL, depending on an individual patient's characteristics of the primary tumor and the location of clinically diagnosed lymph node metastases. As such, the data may contribute to further personalize the elective treatment of the neck for HNSCC patients, i.e. definition of the elective clinical target volume (CTV-N) in radiotherapy (RT) and the extent of neck dissection (ND) in surgery. There exists only one similar publicly available dataset that reports clinical involvement per LNL in 287 oropharyngeal SCC patients [2]. The data presented in this article substantially extends the available data, it additionally includes pathologically assessed involvement per LNL, and it provides data for multiple subsites in the head and neck region.

Synthetic computed tomography (sCT) scans are necessary for dose calculation in magnetic resonance (MR)-only radiotherapy. While deep learning (DL) has shown remarkable performance in generating sCT scans from MR images, research has predominantly focused on high-field MR images. This study presents the first implementation of a DL model for sCT generation in head-and-neck (HN) cancer using low-field MR images. Specifically, the use of vision transformers (ViTs) was explored. The dataset consisted of 31 patients, resulting in 196 pairs of deformably-registered computed tomography (dCT) and MR scans. The latter were obtained using a balanced steady-state precession sequence on a 0.35T scanner. Residual ViTs were trained on 2D axial, sagittal, and coronal slices, respectively, and the final sCTs were generated by averaging the models’ outputs. Different image similarity metrics, dose volume histogram (DVH) deviations, and gamma analyses were computed on the test set (n = 6). The overlap between auto-contours on sCT scans and manual contours on MR images was evaluated for different organs-at-risk using the Dice score. The median [range] value of the test mean absolute error was 57 [37–74] HU. DVH deviations were below 1% for all structures. The median gamma passing rates exceeded 94% in the 2%/2mm analysis (threshold = 90%). The median Dice scores were above 0.7 for all organs-at-risk. The clinical applicability of DL-based sCT generation from low-field MR images in HN cancer was proved. High sCT-dCT similarity and dose metric accuracy were achieved, and sCT suitability for organs-at-risk auto-delineation was shown.