BrainLesion Suite: Modular Lesion Analysis

• BrainLesion Suite is a modular framework for brain lesion analysis that organizes preprocessing, synthesis, segmentation, and evaluation into distinct Python packages.
• It leverages state-of-the-art methods such as affine registration, U-Net architectures, and diffusion models to improve accuracy and reproducibility.
• The suite supports flexible integration via APIs and CLI commands, enabling third-party segmentation and statistical mapping tools to be seamlessly incorporated.

BrainLesion Suite denotes a modular framework for brain lesion image analysis in Python whose core design combines adaptable preprocessing, pathology-specific modeling, and quantitative evaluation under a “brainless” interface intended to minimize workflow overhead (Kofler et al., 11 Jul 2025). In related literature, the same name is also used for a five-stage diffusion-based pipeline that performs abnormal-region segmentation, tissue deformation estimation, deformation reversal, core-lesion inpainting, and healthy-brain estimation, indicating that the term spans both a software toolkit and a broader systems concept for lesion analysis (Zamzam et al., 8 Jul 2025).

1. Core conception and package organization

BrainLesion Suite is described as a versatile toolkit for building modular brain lesion image analysis pipelines in Python, organized as a collection of small, single-purpose Python packages under the GitHub organization https://github.com/BrainLesion/ (Kofler et al., 11 Jul 2025). Its stated design principles are explicitly Pythonic: idiomatic Python practices, PEP8 conventions, minimal boilerplate, duck typing, and a “brainless” interface in which common workflows execute with a single function call or CLI command. The framework was originally developed for image analysis pipelines of brain lesions such as glioma, metastasis, and multiple sclerosis, but it is stated to be adaptable to other biomedical image analysis applications (Kofler et al., 11 Jul 2025).

The package decomposition is central to the framework’s identity.

Package	Role
bls-modsort	interactive GUI for sequence sorting
bls-preprocessing	co-registration, atlas alignment, skull-strip, defacing, bias correction
bls-brats	orchestrator for synthesis/inpainting/segmentation models from the BraTS challenge
bls-aurora	metastasis-specific segmentation
bls-gliomoda	adult glioma segmentation (nnU-Net based)
bls-petu	pediatric tumor segmentation
bls-dqe	deep quality estimation surrogate for human QC
bls-panoptica	instance-wise evaluation metrics

Interoperability is handled through shared data types such as ImageSet and SegmentationMap in a common bls_core helper library, while each package declares CLI commands and Python API objects via setuptools entry points (Kofler et al., 11 Jul 2025). A recurring misconception is to treat BrainLesion Suite as a single segmentation network. The package structure shows the opposite: it is a modular orchestration layer in which preprocessing, synthesis, segmentation, QC, and evaluation are separable components.

2. Preprocessing, registration, and image standardization

The preprocessing module, bls-preprocessing (BLP), is the structural backbone of the toolkit (Kofler et al., 11 Jul 2025). Its multi-modal co-registration pipeline is explicitly defined in five steps: rigid or affine registration of all moving modalities $I_m$ to a chosen center modality $I_c$; rigid or affine+$\mathrm{SyN}$ registration of $I_c$ to a target atlas $A$; application of the $I_c \to A$ transform to all other modalities; optional refinement of all $I_m \to I_c$ transforms again in atlas space; and then skull-strip, deface, N4 bias correction, and intensity normalization.

Affine registration is written as

$$T^* = \arg\min_{T\in \mathrm{Aff}} \mathcal{L}\bigl(I_{\rm fixed},\, I_{\rm moved}\circ T\bigr),$$

with $\mathcal{L}$ typically Mutual Information or Mean Squared Error (Kofler et al., 11 Jul 2025). The paper provides both forms: $$\mathcal{L}_{\rm MSE}=\int \bigl[I_f(x)-I_m(Tx)\bigr]^2\,dx$$ and

$I_c$0

Atlas registration supports ANTsPy or niftyreg affine registration on the linear side, and ANTs SyN, elastix B-splines, or Greedy SyN on the non-linear side (Kofler et al., 11 Jul 2025). Skull-stripping is implemented via HD-BET, defacing via Quickshear, and bias-field correction via N4ITK through SimpleITK, with corrected intensities written as

$I_c$1

Normalization strategies include z-score, histogram matching, and WhiteStripe (Kofler et al., 11 Jul 2025).

These details matter because nearly every downstream module assumes standardized spatial coordinates, harmonized intensities, and optional privacy-preserving defacing. The architecture therefore treats preprocessing not as an auxiliary convenience, but as the condition of possibility for modular downstream substitution.

3. Generative restoration and pathology-specific segmentation

Beyond registration, BrainLesion Suite includes generative and restorative components through bls-brats (Kofler et al., 11 Jul 2025). For missing-modality synthesis, the orchestrator includes top-performing models from the BraTS synthesis task. The typical architecture is a 2D or 3D U-Net generator $I_c$2 mapping available sequences $I_c$3 to a missing modality, together with a PatchGAN discriminator $I_c$4. The training objective combines an $I_c$5 reconstruction term and an adversarial term: $I_c$6 with total loss $I_c$7 (Kofler et al., 11 Jul 2025). The stated protocol is 200 epochs, batch size 4, learning rate $I_c$8, and decay after 100 epochs.

For lesion inpainting, the framework uses a CNN with partial convolutions or a masked U-Net to reconstruct healthy-appearing tissue in lesion cavities (Kofler et al., 11 Jul 2025). The objective is

$I_c$9

where $\mathrm{SyN}$0 is the binary lesion mask. This gives the suite a restorative capability rather than a purely discriminative one.

Pathology-specific tumor segmentation is organized as separate modules for metastasis, adult glioma, and pediatric tumor (Kofler et al., 11 Jul 2025). The core model family is U-Net or nnU-Net with skip connections, deep supervision, and group normalization. Input channels range from $\mathrm{SyN}$1 to $\mathrm{SyN}$2 depending on available modalities, and all channels may be stacked as $\mathrm{SyN}$3; missing channels are filled with synthetics from the synthesis module or zeroed (Kofler et al., 11 Jul 2025). The principal loss is Dice, optionally combined with cross-entropy: $\mathrm{SyN}$4

This design implies that BrainLesion Suite is not limited to lesion detection. It encompasses synthesis, restoration, and segmentation in a unified pipeline, with modality-adaptive behavior built into the package interfaces.

4. Evaluation, quality control, and workflow composition

The evaluation layer is implemented primarily through bls-panoptica, which extracts instances from semantic maps and computes lesion-wise metrics including PQ, SQ, RQ, ASSD, and clDSC (Kofler et al., 11 Jul 2025). A predicted lesion is a true positive when matched to a ground-truth lesion with $\mathrm{SyN}$5; otherwise predicted lesions may contribute false positives and unmatched ground-truth lesions false negatives. Precision and recall are defined in the usual lesion-wise form: $\mathrm{SyN}$6 Panoptic Quality is decomposed as

$\mathrm{SyN}$7

Metrics may be computed per subject and then averaged across the cohort or reported as median $\mathrm{SyN}$8 IQR, with optional stratification by lesion size, region, or modality (Kofler et al., 11 Jul 2025).

Quality control is additionally addressed by bls-dqe, described as a deep quality estimation surrogate for human QC (Kofler et al., 11 Jul 2025). At the workflow level, the suite explicitly supports a staged pipeline from sequence sorting and preprocessing to synthesis, segmentation, quality estimation, and instance-wise evaluation. The provided examples show that these stages can be invoked through either Python APIs or CLI commands, reinforcing the package-level separation of concerns (Kofler et al., 11 Jul 2025).

A plausible implication is that BrainLesion Suite’s principal scientific value lies in reproducible pipeline assembly. Rather than fixing a single model, it standardizes how models are invoked, compared, and audited.

5. Integration of external segmentation and report-supervision methods

The BrainLesion Suite literature repeatedly describes how third-party methods can be encapsulated as callable modules. A concrete example is SPiN, the “Subpixel Network” for small lesion segmentation in brain MRIs (Wong et al., 2021). SPiN uses a standard encoder-decoder augmented by a subpixel embedding branch and a learnable downsampler, with input $\mathrm{SyN}$9, $I_c$0 adjacent axial slices in the paper, four downsampling stages, a bottleneck of two $I_c$1 convolutions with 256 filters, and a decoder whose final subpixel logits are produced at $I_c$2 resolution (Wong et al., 2021). The proposed BrainLesion Suite packaging defines a SubpixelSegmenter with inputs volume, slice_count, threshold, and device, and outputs mask and optional softmap, intended to slot in immediately after image preprocessing such as bias-field correction and skull-stripping (Wong et al., 2021). On ATLAS release 1.1, SPiN is reported to achieve DSC $I_c$3, IOU $I_c$4, Precision $I_c$5, and Recall $I_c$6 on the full held-out test split, with approximately $I_c$7 M parameters, inference memory $I_c$8 GB, and runtime approximately $I_c$9 ms per 2D slice (Wong et al., 2021).

A second integration pattern is MS-RSuper, a report-supervised framework for multimodal MRI tumor segmentation with substructures (Ge et al., 24 Feb 2026). Its decomposition into a Report Parsing Service, Modality–Substructure Mapper, Loss Factory, Segmentation Engine, and Data Manager is explicitly presented as a BrainLesion Suite-compatible design (Ge et al., 24 Feb 2026). The segmentation backbone is a 3D nnU-Net trained on a small fully labeled set $A$0 and a larger report-only set $A$1, with total loss

$A$2

On a held-out test of 50 MEN and 50 MET scans, MS-RSuper improves Dice over both a masks-only baseline and a naive R-Super baseline across WT, TC, and ET; for example, in MET, WT Dice is $A$3 for MS-RSuper versus $A$4 for Masks-Only and $A$5 for R-Super baseline (Ge et al., 24 Feb 2026). Deployment considerations include Docker/Kubernetes containerization, ONNX or TorchScript export, metadata persistence, and HIPAA-compliant isolation of the LLM parsing service (Ge et al., 24 Feb 2026).

A third example is SYNAPSE-Net, presented as a unified framework for heterogeneous brain lesion segmentation and accompanied by explicit BrainLesion Suite integration guidelines (Hassan et al., 30 Oct 2025). Its multi-stream encoder accepts $A$6 modalities, unused streams can be zero-filled or omitted, and inference uses sliding-window axial stripes with Gaussian weighting and 50% overlap, followed by validation-set grid search over threshold $A$7 and minimum lesion size $A$8 voxels (Hassan et al., 30 Oct 2025). Reported results include DSC $A$9 and HD95 $I_c \to A$0 mm on MICCAI 2017 WMH, HD95 $I_c \to A$1 mm on ISLES 2022, and TC DSC $I_c \to A$2 on BraTS 2020 (Hassan et al., 30 Oct 2025).

Taken together, these integrations suggest that BrainLesion Suite functions as a host framework for heterogeneous methodological paradigms: 2.5D small-lesion segmentation, report-supervised 3D nnU-Net training, and hybrid CNN-transformer multi-modal segmentation.

6. Diffusion-based lesion reversal and Bayesian lesion mapping

The term BrainLesion Suite is used in a second, distinct sense in the diffusion-model paper “Modeling and Reversing Brain Lesions Using Diffusion Models” (Zamzam et al., 8 Jul 2025). There, the suite consists of five major stages: abnormal-region segmentation, tissue deformation estimation, deformation reversal, core-lesion isolation, and healthy brain estimation. The forward lesion-growth model uses a registration network $I_c \to A$3 that outputs a dense velocity field $I_c \to A$4, whose integration yields a diffeomorphic displacement $I_c \to A$5, trained with

$I_c \to A$6

with $I_c \to A$7, together with an incompressible-fluid regularizer $I_c \to A$8 (Zamzam et al., 8 Jul 2025). The reverse stage adopts a variance-preserving diffusion process with $I_c \to A$9 timesteps and learns $I_m \to I_c$0 under the standard $I_m \to I_c$1 noise-prediction loss. Reported lesion-segmentation Dice on BraTS21 is $I_m \to I_c$2 for the Jacobian-threshold method, compared with $I_m \to I_c$3 for 2D diffusion, $I_m \to I_c$4 for 3D latent, and $I_m \to I_c$5 for 3D diffusion; in lesioned-brain labeling, Dice increases by approximately $I_m \to I_c$6–$I_m \to I_c$7 overall and approximately $I_m \to I_c$8–$I_m \to I_c$9 within 10 voxels of the lesion (Zamzam et al., 8 Jul 2025). This use of the name broadens the concept of a lesion-analysis suite from segmentation and evaluation to inverse modeling of deformation and pre-lesion anatomy.

A complementary extension is BLESS, a scalable hierarchical Bayesian spatial model for binary lesion images with a structured spike-and-slab prior (Menacher et al., 2022). BLESS models voxelwise lesion indicators $$T^* = \arg\min_{T\in \mathrm{Aff}} \mathcal{L}\bigl(I_{\rm fixed},\, I_{\rm moved}\circ T\bigr),$$0 through probit augmentation, spatially varying regression coefficients, inclusion indicators $$T^* = \arg\min_{T\in \mathrm{Aff}} \mathcal{L}\bigl(I_{\rm fixed},\, I_{\rm moved}\circ T\bigr),$$1, and a multivariate conditional autoregressive prior on the inclusion log-odds $$T^* = \arg\min_{T\in \mathrm{Aff}} \mathcal{L}\bigl(I_{\rm fixed},\, I_{\rm moved}\circ T\bigr),$$2 (Menacher et al., 2022). Mean-field variational inference is combined with dynamic posterior exploration through a decreasing grid of spike variances $$T^* = \arg\min_{T\in \mathrm{Aff}} \mathcal{L}\bigl(I_{\rm fixed},\, I_{\rm moved}\circ T\bigr),$$3, and uncertainty recovery is handled by BB-BLESS through weighted/Dirichlet bootstraps and jittered priors (Menacher et al., 2022). The method also defines cluster-size imaging statistics, including credible intervals of cluster size and voxelwise reliability maps of cluster occurrence. Its stated computational regime reaches $$T^* = \arg\min_{T\in \mathrm{Aff}} \mathcal{L}\bigl(I_{\rm fixed},\, I_{\rm moved}\circ T\bigr),$$4 subjects and $$T^* = \arg\min_{T\in \mathrm{Aff}} \mathcal{L}\bigl(I_{\rm fixed},\, I_{\rm moved}\circ T\bigr),$$5 voxels, with a recommended modular implementation split into Data-Loader Module, Inference Engine, and Post-Processing (Menacher et al., 2022).

This dual extension is significant. One branch pushes BrainLesion Suite toward biomechanical lesion reversal and healthy-brain estimation; the other pushes it toward uncertainty-aware population-level lesion mapping. The shared theme is not a single algorithmic family, but a modular infrastructure for lesion-centric neuroimaging analysis across preprocessing, prediction, restoration, and statistical inference.