Raidionics: CNS Tumor Segmentation Platform

• Raidionics is an open-source software platform that automates central nervous system tumor segmentation and standardized report generation from MRI scans.
• It integrates deep-learning models with versatile deployment modes—including a stand-alone desktop app and a 3D Slicer plugin—to enhance clinical workflows.
• The platform employs advanced preprocessing, atlas registration, and unified FLAIR analysis while addressing challenges in small-tumor and postoperative segmentation.

Raidionics is an open-source software platform for automatic central nervous system tumor segmentation from MRI and for generation of standardized clinical reports derived from those segmentations. It was introduced as a practical translation of trained deep-learning models for preoperative brain tumor imaging, packaged for direct use by clinicians and researchers, and subsequently expanded into a broader platform covering preoperative and early postoperative workflows, multiple deployment modes, and later a unified FLAIR hyperintensity segmentation capability integrated into the same ecosystem. Across the cited studies, Raidionics appears in two principal user-facing forms: a stand-alone desktop application and Raidionics-Slicer, a plugin integrated into 3D Slicer (Bouget et al., 2022, Bouget et al., 2023, Faanes et al., 19 Dec 2025).

1. Development trajectory and scope

Raidionics emerged in response to a specific workflow bottleneck in neuro-oncology: multidisciplinary decisions depend on preoperative MR imaging, yet standardized and automatic methods for tumor delineation and structured image-based reporting were lacking. The 2022 study framed the software around four major tumor groups—glioblastomas, lower grade gliomas, meningiomas, and metastases—and coupled automatic segmentation with standardized report generation. The 2023 study broadened the framing from preoperative brain tumor imaging to pre- and postoperative central nervous system tumor analysis, added an early postoperative glioblastoma residual tumor segmentation model, and formalized the software as an open-source platform with a graphical user interface and a stable backend. The 2025 study extended the platform further by integrating a unified FLAIR hyperintensity model for multiple tumor types and acquisition time points, including pre-operative and post-operative scans (Bouget et al., 2022, Bouget et al., 2023, Faanes et al., 19 Dec 2025).

Study	Contribution to Raidionics	Scope emphasized
"Preoperative brain tumor imaging: models and software for segmentation and standardized reporting" (Bouget et al., 2022)	Introduction of Raidionics and Raidionics-Slicer	Preoperative segmentation and reporting
"Raidionics: an open software for pre- and postoperative central nervous system tumor segmentation and standardized reporting" (Bouget et al., 2023)	Open software platform with GUI, backend, and postoperative module	Preoperative plus early postoperative GBM
"A unified FLAIR hyperintensity segmentation model for various CNS tumor types and acquisition time points" (Faanes et al., 19 Dec 2025)	Unified FH model integrated into Raidionics	Cross-tumor, pre-/post-operative FLAIR analysis

The software’s intended users are consistently described as clinicians without programming experience, neuroradiologists, neurosurgeons, oncologists, multidisciplinary brain tumor teams, and researchers who need rapid generation of annotated datasets or standardized quantitative measurements. The authors explicitly position the system against manual delineation, visual estimation, and fragmented deployment of research models. In that sense, Raidionics is not only a segmentation package but also a translational layer between model development and routine image-analysis workflows (Bouget et al., 2022, Bouget et al., 2023).

2. Software architecture, deployment model, and data handling

The 2023 software paper describes Raidionics as a user-facing application coupled to a backend processing engine. The graphical interface exposes two operating modes. In single-patient mode, the interface is divided into a left panel for patient import and report browsing, a center panel with linked axial, coronal, and sagittal viewers synchronized in 3D space, and a right panel listing MR scans, annotations, and atlases, with configurable overlay color and opacity and an “Actions” tab for segmentation or standardized reporting. In batch or study mode, the interface supports cohort import, study-level launch of segmentation and reporting, and summary tables for file content, annotations, and extracted reporting parameters (Bouget et al., 2023).

The backend is also usable independently through a command-line interface, as a Python library, and as a Docker container. A 3D Slicer plugin is available and uses the backend Docker container. For inference, the platform uses ONNX Runtime, which the authors selected for interoperability across frameworks and operating systems. Prebuilt installers are reported for Windows 10 or newer, Ubuntu Linux 18.04 or newer, and macOS Catalina 10.15 or newer, including ARM-based Apple M1 systems (Bouget et al., 2023).

Raidionics supports data import from raw DICOM folders and from converted volumes in NIfTI (.nii, .nii.gz), MetaImage (.mhd), and NRRD (.nrrd), as well as more generally any format supported by SimpleITK. All MR scans are internally converted to NIfTI for processing. External annotation files can also be imported, provided they match the corresponding MR scan in shape, spacing, and orientation; if multiple scans are present, the annotation must be linked to the correct scan through a “Parent MRI” selector. The software expects MR scans to be associated with timestamps in ascending order so that preoperative and postoperative data can be distinguished when loaded for the same patient. Results are stored in the patient folder together with a custom .raidionics scene file; outputs include NIfTI for images, segmentations, and atlases, CSV for statistics, and standardized reports in TXT, JSON, and CSV formats (Bouget et al., 2023).

3. Segmentation tasks, model families, and preprocessing pipelines

The original preoperative Raidionics workflow is tumor-type specific and sequence specific. In the 2022 study, each tumor type is associated with a single MRI sequence per patient: glioblastoma uses gadolinium-enhanced T1-weighted MRI ($T1c$), lower grade glioma uses FLAIR, meningioma uses $T1c$, and metastasis uses $T1c$. The user loads a scan, manually selects the tumor type, and can either run automatic segmentation or provide a pre-existing tumor mask to bypass segmentation. The software then registers the tumor mask to MNI space, computes tumor-derived quantitative and spatial features, and generates a standardized clinical report. The requirement that tumor type be entered manually is explicitly identified as a current limitation, with automatic classification described as future work (Bouget et al., 2022).

The 2022 segmentation models are based on AGU-Net. The reported preprocessing pipeline comprises resampling to isotropic $1\,\mathrm{mm}^3$ spacing using NiBabel spline interpolation of order 1, tumor-type-specific use of either tight clipping around the patient’s head or skull-stripping using a custom brain segmentation model, resizing to $128 \times 128 \times 144$ voxels, and intensity normalization to $[0,1]$. Tumor-type-specific training choices are also reported: glioblastoma used skull-stripping and leave-one-hospital-out evaluation; lower grade glioma used tight clipping and 5-fold cross-validation; meningioma used tight clipping and 5-fold cross-validation; metastasis used skull-stripping, transfer learning, and 5-fold cross-validation. The training framework is specified as TensorFlow v1.13.1 with Python 3.6, Adam optimization, batch size 32 samples with accumulated gradients and actual batch size 2, early stopping after 30 consecutive epochs without validation loss improvement, and an initial learning rate of $1e^{-3}$ for training from scratch or $1e^{-4}$ for transfer learning; the metastasis model was fine-tuned from the best glioblastoma model (Bouget et al., 2022).

The 2023 software paper describes the deployed preoperative models as AGU-Net variants with five levels, filter sizes $\{16, 32, 128, 256, 256\}$, deep supervision, multiscale input, and single attention modules. Relative to the originally published architecture, the deployed version removes all batch normalization layers and uses a patch-wise strategy for training and inference with patch size $160^3$ voxels. Its preprocessing pipeline is reported somewhat differently from the 2022 model-development paper: isotropic resampling to $T1c$0, skull stripping except for meningioma, intensity clipping to remove the top $T1c$1 highest values, and intensity normalization and scaling to $T1c$2. The postoperative module is an early postoperative glioblastoma residual tumor segmentation model adopted from a related study; the 2023 paper packages it into the software but does not restate its internal architecture (Bouget et al., 2023).

The 2025 extension adds a unified FLAIR hyperintensity segmentation capability. This model operates on a single FLAIR MRI scan and is intended to segment FH across multiple tumor categories and multiple acquisition time points, including pre-operative meningiomas, pre-operative metastases, pre-operative and post-operative gliomas from BraTS, and pre-operative and post-operative lower-grade gliomas from a private dataset. The paper states that the model is integrated into Raidionics. Its preprocessing pipeline consists of resampling to isotropic $T1c$3 spacing using first-order spline interpolation, tight cropping around the patient’s head, intensity clipping to the $T1c$4 percentile range, and zero-mean normalization over nonzero voxels. Inference uses sliding-window prediction with patch size $T1c$5 and $T1c$6 overlap. Postprocessing applies a binary brain mask, removes predictions with volume below $T1c$7, and removes predictions not visible in two consecutive 2D slices. Architecturally, all models in that study use Attention U-Net with 5 levels, filter sizes $T1c$8, instance normalization, dropout $T1c$9, and input size $T1c$0 (Faanes et al., 19 Dec 2025).

4. Standardized reporting, atlas registration, and derived quantitative descriptors

Standardized reporting is a core design objective of Raidionics rather than a post hoc add-on. In the 2022 formulation, the software pipeline explicitly combines automatic tumor segmentation, atlas-based spatial analysis, tumor feature extraction, and standardized report generation. In the 2023 system description, report generation is tied to registration of the patient MRI to the symmetric MNI ICBM2009a atlas using the SyN method from ANTs, enabling atlas-based computation of spatial descriptors (Bouget et al., 2022, Bouget et al., 2023).

For preoperative reporting, the 2023 paper states that the extracted feature set includes tumor volume, laterality, multifocality, cortical structure location profile, and subcortical structure location profile. For glioblastoma only, the report also includes resectability features; the paper explicitly notes that these are not available for the other tumor types. The resulting report is stored in text, JSON, and CSV forms. The 2022 paper presents the same general workflow from the clinical perspective: the user may either accept automatic segmentation or supply a pre-existing tumor mask, after which the mask is registered to MNI space and tumor-derived quantitative and spatial features are computed for a reproducible summary of tumor characteristics that would otherwise often be estimated by visual inspection (Bouget et al., 2022, Bouget et al., 2023).

For postoperative reporting, the scope is narrower and focused on early postoperative glioblastoma assessment. The report automatically extracts preoperative volume, postoperative residual volume, extent of resection (EOR), and EOR-based patient classification. The 2023 paper remarks that the postoperative report currently limits itself mainly to the extent of resection. A faithful reconstruction of the volumetric definition consistent with the manuscript is

$T1c$1

This suggests that postoperative Raidionics is organized around residual-disease quantification rather than a full radiological characterization of postsurgical anatomy (Bouget et al., 2023).

The 2025 FH extension broadens the space of reportable biomarkers. The primary target is 3D segmentation of FLAIR hyperintensity volume from a single FLAIR scan. For BraTS-derived group B data, the study distinguishes FH from surrounding non-enhancing FLAIR hyperintensity (SNFH), with SNFH obtained by subtracting tumor volumes from FH predictions. The paper explicitly notes that, for enhancing tumors such as meningioma, glioblastoma, and metastasis, tumor core volume is often already known from manual delineation or from “high-performing segmentation software, such as Raidionics.” This makes the FH model a natural extension of the existing Raidionics ecosystem rather than an isolated module, because it can be combined with existing tumor segmentations to derive edema- or infiltration-related quantities (Faanes et al., 19 Dec 2025).

5. Validation results, runtime characteristics, and operational envelope

The 2022 study reports that segmentation performances were “quite homogeneous” across the four major preoperative brain tumor types, with an average true positive Dice ranging between $T1c$2 and $T1c$3, patient-wise recall between $T1c$4 and $T1c$5, and patient-wise precision around $T1c$6. The same paper reports that, on a desktop computer with CPU support, tumor segmentation can be performed in 16 to 54 seconds depending on MRI volume dimensions, while generation of a standardized clinical report including tumor segmentation and feature computation requires 5 to 15 minutes. Its cohort summary comprises 2134 glioblastoma cases, 659 lower grade glioma cases, 719 meningioma cases, and 396 metastasis cases, totaling about 3908 patients (Bouget et al., 2022).

The 2023 software paper reports a related but not identical validation set composition: 2125 glioblastoma scans from 15 institutions, 678 lower-grade glioma scans from 4 institutions, 706 meningioma scans from 2 institutions, and 394 metastasis scans from 2 institutions. Across the preoperative models, the headline result is average Dice around $T1c$7 with patient-wise recall and precision around $T1c$8. By tumor type, the reported voxel-wise Dice values are $T1c$9 for preoperative glioblastoma, $1\,\mathrm{mm}^3$0 for lower-grade glioma, $1\,\mathrm{mm}^3$1 for meningioma, and $1\,\mathrm{mm}^3$2 for metastasis. Postoperatively, the early postoperative glioblastoma residual tumor model is markedly weaker, with voxel-wise Dice $1\,\mathrm{mm}^3$3. The same paper states that a standardized clinical report, including segmentation and feature computation, requires about ten minutes on a regular laptop (Bouget et al., 2023).

The 2023 paper also provides hardware-stratified runtime measurements. On the desktop system, total processing times are $1\,\mathrm{mm}^3$4 minutes for meningioma, $1\,\mathrm{mm}^3$5 minutes for metastasis, $1\,\mathrm{mm}^3$6 minutes for lower-grade glioma, $1\,\mathrm{mm}^3$7 minutes for preoperative glioblastoma, and $1\,\mathrm{mm}^3$8 minutes for postoperative glioblastoma. On the laptop system, the corresponding totals are $1\,\mathrm{mm}^3$9, $128 \times 128 \times 144$0, $128 \times 128 \times 144$1, $128 \times 128 \times 144$2, and $128 \times 128 \times 144$3 minutes. The paper attributes longer preoperative reporting times mainly to registration and notes that very high-resolution images markedly increase registration time (Bouget et al., 2023).

The 2025 FH study reports cohort-level validation results for the unified FLAIR model now integrated into Raidionics. For FH segmentation with the unified model, the reported Dice scores are $128 \times 128 \times 144$4 for pre-operative meningiomas, $128 \times 128 \times 144$5 for pre-operative metastases, $128 \times 128 \times 144$6 for pre-operative gliomas from BraTS, $128 \times 128 \times 144$7 for post-operative gliomas from BraTS, $128 \times 128 \times 144$8 for pre-operative lower-grade gliomas, and $128 \times 128 \times 144$9 for post-operative lower-grade gliomas. Detection rates range from $[0,1]$0 in post-operative lower-grade gliomas to $[0,1]$1 in pre-operative BraTS gliomas. The same study reports that the unified model achieved comparable performance to dataset-specific models on their respective datasets and, for SNFH after tumor subtraction, even exceeded the dataset-specific SNFH model on BraTS pre-operative and post-operative data. The paper also benchmarks the unified model against BraTS leaderboards and notes that the Raidionics-integrated model reaches BraTS-comparable performance despite using only one MRI sequence rather than four (Faanes et al., 19 Dec 2025).

Taken together, these results define a clear operational envelope. Preoperative segmentation is the strongest and most mature part of the platform; unified FLAIR analysis is robust across heterogeneous tumor types and time points; early postoperative glioblastoma residual segmentation and early postoperative FLAIR interpretation remain substantially more difficult. This suggests that Raidionics is best aligned with quantitative support and standardized reporting in routine preoperative analysis and in later-stage follow-up, while immediate postsurgical use cases require greater caution (Bouget et al., 2023, Faanes et al., 19 Dec 2025).

6. Limitations, failure modes, and projected evolution

Several limitations are stated explicitly across the papers. In the 2022 system, tumor type selection must be entered manually by the user, and the authors identify automatic classification of the brain tumor type as future work. The same paper also notes that inclusion of postoperative segmentation in both software solutions will be necessary for complete postoperative standardized clinical reports, indicating that the original preoperative system did not yet cover the full perioperative trajectory (Bouget et al., 2022).

The 2023 software paper emphasizes two persistent technical weaknesses: small-tumor performance and postoperative segmentation accuracy. For meningioma, metastasis, and glioblastoma, tumors below $[0,1]$2 are substantially harder; for lower-grade glioma, the analogous cutoff used in the analysis is $[0,1]$3. The authors conclude that performance on non-small tumors is excellent, whereas performance for small tumors is substantially lower, with average Dice closer to $[0,1]$4 and patient-wise recall around $[0,1]$5, which limits current use for early-stage detection or screening. The postoperative glioblastoma module is also much less accurate than the preoperative models, and the paper explicitly states that better postoperative segmentation models still need to be investigated. In addition, the authors note that current reports may not yet be exhaustive enough for full preoperative surgical assessment, that the postoperative report is deliberately focused mainly on EOR, and that the development datasets cannot be publicly released because of GDPR restrictions (Bouget et al., 2023).

The 2025 FLAIR extension introduces a different class of limitation: annotation heterogeneity across sources. The paper states that there was no common annotation protocol across all datasets; group A lower-grade glioma labels were tumor masks repurposed as FH labels, whereas group B FH labels were formed by merging tumor and SNFH masks. Multiple annotators and centers contributed, multifocal disease could be treated inconsistently, and post-operative lower-grade glioma ground truth may underestimate true FH because edema can be present while only residual tumor tissue was annotated. The authors additionally note that small lesions, especially below approximately $[0,1]$6, remain difficult; early post-operative scans are particularly challenging; using only FLAIR may be suboptimal where T2 is preferred; and more clinical validation is needed before full clinical implementation (Faanes et al., 19 Dec 2025).

Future directions are correspondingly diverse. The 2022 paper points to automatic tumor-type classification and broader postoperative support. The 2023 paper lists more segmentation models, more report features, better postoperative models, segmentation of postoperative complications such as hemorrhage or infarction, a metrics computation module, a heatmap location generator, and additional structural segmentation models. The 2025 paper proposes standardized relabeling or active learning, more true negative cases, improved handling of small structures, multiple MRI sequences as input, multi-class segmentation of FH, contrast enhancement, and cavity, and synthetic MRI generation for missing-sequence completion or conversion of thick-slice scans into more usable 3D representations (Bouget et al., 2022, Bouget et al., 2023, Faanes et al., 19 Dec 2025).

In aggregate, Raidionics can be understood as an evolving neuro-oncology image-analysis platform with three tightly coupled functions: deployment of trained segmentation models, generation of standardized atlas-aware quantitative reports, and progressive extension from preoperative tumor-core analysis toward postoperative and FLAIR-based longitudinal characterization. Its principal significance lies not only in segmentation accuracy, but in the software-level integration of model inference, anatomical registration, feature extraction, and exportable reporting into a reproducible open workflow for CNS tumor analysis (Bouget et al., 2022, Bouget et al., 2023, Faanes et al., 19 Dec 2025).