01 · Input and data

Input & Data

Complete or incomplete multi-sequence MRI can be interpreted with optional patient metadata.

T1T1cT2T2-FLAIRAge · sex · text

What can BrainVLM receive?

The model accepts the heterogeneous inputs that appear in routine neuro-oncology workflows.

Flexible input

Complete or incomplete MRI studies

Use a full multi-parametric study or submit the sequences available for a patient. Demographics and available radiology text can be added as complementary clinical context.

Full protocolT1 · T1c · T2 · FLAIR

Partial protocolMissing-sequence tolerant

Image representationRaw or pre-processed MRI

MetadataAge · sex · clinical text

BrainVLM multi-sequence MRI input and case-level interpretation — Multi-sequence input: T1, T1c, T2, and FLAIR can be interpreted as one case, with raw or pre-processed images and metadata.

Multi-sequence evidence, not single-sequence guessing

The same lesion can look different across MRI sequences, and each view contributes a distinct part of the diagnostic picture. BrainVLM integrates these complementary signals at the study level before producing a case-level diagnosis, report, and confidence estimate.

Representative multimodal MRI case: anatomical appearance, enhancement, internal signal, and lesion extent are interpreted together rather than inferred from a single sequence.

Complementary evidence

Each sequence answers a different question

T1Anatomy and baseline lesion signal

T1cContrast enhancement and blood–brain barrier disruption

T2Fluid-rich tissue and internal lesion characteristics

T2-FLAIREdema, infiltration, and periventricular abnormality

Study-level integration

Available sequences and views are combined dynamically

BrainVLM selects complementary images from the sequences and anatomical views available for each patient. This lets it preserve case-level context when a protocol is incomplete, while still benefiting from the richer evidence provided by a complete multiparametric examination.

Missing T1, T2, or FLAIR<5% overall F1 reduction

Missing T1c13.4% F1 reduction

InterpretationSequence value varies by tumor type

Missing-sequence tolerance does not mean that all sequences are interchangeable. Ablation and Shapley-value analyses both identified T1c as the most influential sequence overall, while the contribution of T1, T2, and T2-FLAIR varied across tumor categories. The model is therefore designed to use all available evidence and degrade gracefully when part of the protocol is absent.

Data scale and clinical diversity

BrainTumor48K was assembled to cover the breadth of brain tumor diagnosis rather than only the most common classes.

47,947Individuals in the curated dataset

39Public repositories integrated

12Collaborating medical centers

12Major WHO CNS5 tumor categories