BACKGROUND: Intracranial hemorrhages (ICH) are life-threatening conditions that require rapid detection and precise subtype classification. Automated segmentation and volumetric analysis using deep learning can enhance clinical decision-making. INTRODUCTION: This study aimed to develop and evaluate 3D U-Net based deep learning models for automatic segmentation and classification of ICH and related pathologies in non-contrast brain CT images. We investigated the effect of the number of segmentation classes on model performance. MATERIALS AND METHODS: A dataset of 1508 non-contrast CT series, sourced from our hospital, the QURE500 dataset, and the RSNA 2019 brain hemorrhage dataset, was curated. Manual annotations by experienced radiologists segmented images into brain parenchyma, cerebrospinal fluid, parenchymal edema, pneumocephalus, and various hemorrhage subtypes. Preprocessing involved converting CT series to NIfTI format and reconstructing slice thickness using the TomoPy library. Six separate 3D U-Net models were trained with different label sets, validated using a 5-fold cross-validation approach. Performance metrics, including Dice score, precision, sensitivity, specificity, and negative predictive value, were calculated. RESULTS: The highest ICH detection sensitivity was 99%, with specificity of 82%, precision of 92%, and a negative predictive value of 98%. The best Dice score for ICH was 0.7. False negatives predominantly occurred in hemorrhages smaller than 1 mL. Brain parenchyma and cerebrospinal fluid segmentations achieved Dice scores of 0.99 and 0.95, respectively. Dice scores for hemorrhage subtypes varied, with intraparenchymal hemorrhage (IPH) achieving the highest at 0.61. Inter-operator agreement was high, with Cohen's kappa values ranging from 0.78 to 0.98. CONCLUSION: Our models effectively detect and classify ICH with high sensitivity and specificity, providing valuable volumetric analysis. The inclusion of diverse clinical data contributed to lower Dice scores compared to the literature but offered realistic performance in a clinical setting. These models can significantly aid in patient triage, surgical planning, and improving clinical outcomes. Further studies are needed to optimize their performance for medical imaging tasks.