Computed tomography (CT) measures X-ray attenuation with a detector array, typically of scintillator–photodiode elements, and reconstructs the signals into Hounsfield-unit images, making it the most widely used imaging sensor for assessing acute intracerebral hemorrhage. However, sensor-derived CT images contain quantum noise, partial-volume blurring, and weak density contrast at lesion boundaries, which limit the consistency of manual interpretation. We developed the Residual–Dense–Attention U-Net with a convolutional block attention module (RDAC U-Net), a signal-processing and feature-fusion back end that converts sensor-derived CT data into pixel-level hemorrhage masks and volumetric models. Residual blocks, dense blocks, channel–spatial attention, and attention gates were combined to enhance lesion signals, suppress background noise, and sharpen boundaries. Hyperparameters were optimized by uniform design with a Kriging surrogate model, and fivefold cross-validation was used to evaluate generalizability. For 1340 slices from 100 patients, the optimized RDAC U-Net achieved a mean Dice coefficient of 0.9213 with a standard deviation of 0.0063, outperforming four attention-based U-Net variants while requiring the shortest training time per epoch. Stacking the slice-wise masks yielded three-dimensional hemorrhage models, extending the CT sensing chain from attenuation measurement to quantitative hematoma volumetry. The study is limited by single-center data and slice-based two-dimensional processing; multicenter, multivendor validation and three-dimensional extension will be conducted in future work.