Background and Objective: Segmenting head structures in medical images is challenging due to anatomical heterogeneity and structural complexity. Multi-task learning can improve performance through shared representations and may help address these challenges. However, architectures such as U-Net can struggle with distorted anatomy, while 3D shape-learning approaches often increase computational cost, motivating efficient alternatives.Methods: We propose DeepCTE3D, a multi-task model based on an adapted U-Net architecture, called И-Net, for segmenting the intracranial space, lateral ventricles, and eyeballs from 3D CT images. The model was trained and evaluated on 559 real-world CT scans from newborns to adults, including cases with severe anatomical distortions such as enlarged ventricles, skull deformities, and abnormal eyeballs. We hypothesized that segmentation complexity and morphological heterogeneity affect learning performance and generalization within a multi-task framework.Results: DeepCTE3D achieved Dice similarity coefficients of 98.13% for the intracranial space, 80.76% for the lateral ventricles, and 87.73% for the eyeballs. Because of their greater morphological heterogeneity and segmentation complexity, the lateral ventricles served as a reference task. Compared with single-task models, the multi-task approach improved training convergence while maintaining comparable or superior performance, particularly enhancing generalization for challenging structures.Conclusions: DeepCTE3D effectively segmented complex head structures across diverse CT datasets, including anatomically distorted cases. The findings support the hypothesis that accounting for segmentation complexity and morphological heterogeneity improves model generalization, especially for highly variable structures. These results highlight the potential of multi-task learning to extend segmentation capabilities to additional head structures and support CT imaging applications, particularly in low-resource settings.