Pixel-level annotation of brain MRI across pathologies
Multi-sequence segmentation of infarcts, hemorrhages, tumors and infections on brain MRI following a detailed radiological protocol.
A neuroimaging AI company needed a large training set covering the full spectrum of brain MRI pathologies. Each study included multiple sequences (FLAIR, DWI, ADC, SWI/T2*, T1, T1+Gd, T2) and required staged decisions on image quality, exclusion criteria and per-pathology annotation.
Assess technical quality per sequence, apply exclusion criteria, then segment intra- and extra-axial pathologies on the correct reference sequence: infarcts (acute / subacute / indeterminate / gliosis), ICH, SDH/EDH, SAH, intra-/extra-axial tumors, infections and additional findings.
Radiologists were trained on the client's protocol and RedBrick-based platform, reading the radiology report first to guide sequence selection and label choice. A shared decision tree handled sequence mismatches, discrepancies with the report and multi-label cases (e.g., tumor + hemorrhage + edema).
The client received consistent, protocol-compliant segmentations across a broad pathology mix, with structured handling of quality issues and edge cases. Weekly deliveries kept the labeling pipeline aligned with the model iteration cadence.