STBI Dataset: Serial Brain CT Scans for Traumatic Brain Injury, Including Detailed Image Labels, Clinical Data, and Outcome Metrics

The imaging data are derived from serial non-contrast head CT examinations performed during hospitalization.File structureThe resource is organized by patient, with chronologically ordered serial scans within each patient directory. The release includes:DICOM image filesNIfTI image filesstructured imaging labelstemporal metadatamultimodal clinical/outcome metadata for the clinical subsetMain imaging variablesThe annotations include:AbnormalityMarshall CT scoreSoft tissue swelling or hematoma (with side and site)PneumocephalusHydrocephalusMass effect (with subtypes)HerniationMidline shift (with quantity)Traumatic subarachnoid hemorrhageIntraventricular hemorrhageEpidural hematoma (with side and site)Subdural hematoma (with side and site)Subdural hygroma (with side and site)Contusion (with side and site)Edema (with subtypes)Skull fracture type and locationBasal cistern statusPostoperative findingsNew neurosurgical intervention visible on CTDataset organizationThe dataset is organized at the patient level.

Within each patient folder, serial CT examinations are arranged chronologically to preserve temporal order and support longitudinal analysis. Along with the image files, three CSV files are provided:1. CT-based fileContains examination-level information for each CT study, including the CT sequence, imaging labels, temporal variables, and scan-level data.2.

Patient-based fileContains patient-level imaging summaries derived from the serial examinations, allowing analysis at the individual patient level rather than the scan level.3. Patient-based file with clinical data (multimodal)Contains patient-level clinical, laboratory, treatment, and outcome variables (mortality and Extended Glasgow Outcome Scale) linked to the imaging data for the multimodal subset of 1,259 patients.Temporal variables and event definitionFor each CT study, study date and time were extracted from DICOM metadata.

Two temporal variables were derived: accumulated time from the first CT and the time between consecutive CT scans. A patient-level surgery label and time-to-event variable were also defined. The event time was anchored to the preoperative CT, because that examination is temporally closest to the operative decision.

Patients without surgery were censored at last available CT or hospital discharge.Usage

Notes

This resource is useful for lesion detection, lesion subtype classification, serial progression modeling, outcome prediction, and other machine learning tasks in traumatic brain injury. Because many patients have repeated CT examinations, scans from the same patient are not statistically independent. Users should therefore split the data at the patient level rather than the scan level.

The dataset is intentionally released without a fixed train-validation-test split so that users can define task-specific splits appropriate for their own modeling goals. For convenience during data transfer, the dataset has been randomly divided into multiple parts to simplify the upload and download process. The resource does not include volumetric lesion segmentation.