2.1. Datasets

Four different datasets including the MRI data from a prospective randomized Phase II and III trials in neuro‐oncology (EORTC‐26101) (Wick et al., 2017; Wick et al., 2016) and three independent public datasets (LONI Probabilistic Brain Atlas [LPBA40], Nathan Kline Institute Enhanced Rockland Sample Neurofeedback Study [NFBS], Calgary‐Campinas‐359 [CC‐359]) (Puccio, Pooley, Pellman, Taverna, & Craddock, 2016; Shattuck et al., 2008; Souza et al., 2018) were used for the present study. The characteristics of the individual datasets were as follows.

The EORTC‐26101 study was a prospective randomized Phase II and III trials in patients with first progression of a glioblastoma after standard chemoradiotherapy. Briefly, Phase II trial evaluated the optimal treatment sequence of bevacizumab and lomustine (four treatment arms with single agent vs. sequential vs. combination) (Wick et al., 2016) whereas the subsequent Phase III trial (two treatment arms) compared patients treated with lomustine alone with those receiving a combination of lomustine and bevacizumab (Wick et al., 2017). Overall, the EORTC‐26101 study included n = 596 patients (n = 159 from Phase II and n = 437 from Phase III) with n = 2,593 individual MRI exams acquired at 37 institutions within Europe. The study was conducted in accordance with the Declaration of Helsinki and the protocol was approved by local ethics committees and patients provided written informed consent (EudraCT# 2010‐023218‐30 and {"type":"clinical-trial","attrs":{"text":"NCT01290939","term_id":"NCT01290939"}}NCT01290939). Full study design and outcomes have been published previously (Wick et al., 2016; Wick et al., 2017). MRI exams were acquired at baseline and every 6 weeks until Week 24, afterward every 3 months. For the present analysis, we included T1‐w, contrast‐enhanced T1‐w (cT1‐w), fluid attenuated inversion recovery (FLAIR), and T2‐weighted (T2‐w) sequences (either acquired 3D and/or with axial orientation) and excluded those with heavy motion artifacts or corrupt data. These criteria were fulfilled by n = 10,005 individual sequences (including n = 2,401 T1‐w, n = 2,248 T2‐w, n = 2,835 FLAIR and n = 2,521 cT1‐w sequences from n = 2,401 exams, and n = 583 patients) which were included for the present analysis. The EORTC‐26101 dataset was divided into a training and test set using a random split of the dataset (∼2:1 ratio) with the constraint that all patients from each of the 37 institution were either assigned to the training or test set (to limit the potential of overfitting the HD‐BET algorithm). By applying this split, the EORTC‐26101 training set included data from n = 25 institutions (n = 6,586 individual MRI sequences from n = 1,568 exams, n = 372 patients) whereas the EORTC‐26101 test set included data from the remaining n = 12 institutions (n = 3,419 individual MRI sequences from n = 833 exams, n = 211 patients). In this context, it is important to emphasize that the EORTC‐26101 test set was entirely independent from the training set, as it is comprised of acquisitions from different institutions (and thus different MRI scanners/field strengths, see Table 1 for the T1 detailed information on the individual MRI sequences, scanner types, field strengths) that are disjunct from the institutions in the training set.

We used three public datasets for independent testing. Specifically, we collected and analyzed data from (a) the single‐institutional Laboratory of Neuro Imaging (LONI) (LPBA40) dataset of the LONI consisting of n = 40 MRI scans from individual healthy human subjects (Shattuck et al., 2008), (b) the single‐institutional NFBS dataset consisting of n = 125 MRI scans from individual patients with a variety of clinical and subclinical psychiatric symptoms (Puccio et al., 2016), and (c) the CC‐359 dataset consisting of n = 359 MRI scans from healthy adults (Souza et al., 2018). For each subject, the repositories contains an anonymized (defaced) T1‐w MRI sequence and a manually corrected ground‐truth (GT) brain mask.