TBI-HemoCT Data set

The TBI-HemoCT Dataset is a large-scale, comprehensive collection of approximately 60,000 axial brain CT images, meticulously curated to serve the machine learning community. The data was sourced from non-contrast head CT scans of patients at Shrimann Superspeciality Hospital and was developed to provide a robust resource for building and validating deep learning models for intracranial hemorrhage detection. The dataset is intuitively organized into six primary subfolders based on the main diagnosis: epidural, subdural, subarachnoid, intraparenchymal, intraventricular, and any (normal). To address the common issue of class imbalance and prevent model bias, the underrepresented epidural category has been synthetically augmented using various image transformations. These augmented files are easily identifiable by the aug[random_number] suffix appended to their filenames, ensuring transparency for researchers.

To further enhance accessibility and efficiency, a descriptive naming convention and a comprehensive log file are provided. The file naming structure is self-contained, following a consistent ID_[Number][Which hemorrhage][Details].png format. This convention allows for immediate identification of a file's unique ID, its primary diagnosis (matching its folder location), and specific details about the findings. The final segment indicates UNIQUE if only the primary hemorrhage is present; in the any (normal) folder, a UNIQUE designation confirms the complete absence of any hemorrhage. If multiple hemorrhages co-exist, all present types are listed, separated by hyphens. Complementing this is the detailed log file, final_log_file.csv, which catalogs every image with columns for file (the exact filename), destination (the folder location), label (1 for hemorrhage, 0 for normal), placed in multiple locations (a boolean indicating if the original scan had multiple findings), and all positive locations (a complete list of all hemorrhage types present), providing a complete ground truth for advanced, multi-label classification Tasks.