Improved 3D U-Net Model for Precise Brain Hematoma Segmentation and Volume Measurement

Cerebral hematoma is a common brain disease that can lead to damage and death of brain tissue. Segmentation and volume measurement of cerebral hematoma are important steps in diagnosis and treatment, but current methods rely on manual operation, are time-consuming, and subject to subjective errors. This paper proposes a cerebral hematoma segmentation and volume measurement algorithm based on a 3D U-net attention model, which can automatically segment the cerebral hematoma area from Computed Tomography(CT) images and calculate its volume. The algorithm uses 3D U-net as the basic network structure, enhances the expression ability of features using attention mechanism, and introduces deep supervision and class balance strategy to improve segmentation accuracy and robustness. Additionally, a conditional model is used for prediction of different types of hematomas. The experiment, conducted on a publicly available dataset, yielded promising outcomes. The 3D U-net improved algorithm demonstrated outstanding performance, with an RMSE (Root Mean Square Error) of 3.12 mL and an MAE (Mean Absolute Error) of 2.53 mL. Additionally, it exhibited exceptional segmentation accuracy, boasting an mIOU (mean Intersection over Union) of 0.754 and a Dice coefficient of 0.864.