Hybrid neural network architecture with temporal interphase attention mechanism for segmentation and classification of multiphase CT images

The subject of this article is neural network architectures for semantic segmentation of multiphase computed tomography (CT) images and methods for integrating temporally ordered phase data based on attention mechanisms. The aim of the work is to develop and experimentally validate a hybrid architecture, MultiPhaseFormer, which provides high-accuracy segmentation of tumor lesions and classification of pathology type by explicitly modeling the dynamics of contrast agent accumulation across phases without requiring prior spatial image registration. The article proposes a Temporal Phase Inter-attention (TPIA) mechanism, in which contrast phases are treated as an ordered temporal sequence, and the query of the attention mechanism is formed from the reference portal-venous phase. The models were trained and validated under identical conditions using the publicly available Primary Liver Cancer CECT dataset. Three architectures of increasing complexity were developed: BaselineUNet3D (using early phase fusion), HybridTPIANetLite (based on averaged phase attention), and the target model MultiPhaseFormer (using voxel-wise temporal attention). A multi-task loss function combining Dice loss, Focal Tversky loss, and cross-entropy was employed. The scientific novelty of the work lies in modeling CT phases as an ordered temporal sequence with learnable phase embeddings, and in implementing directed interphase attention with a reference phase, enabling implicit feature alignment at the level of deep representations without computationally expensive 3D registration. According to experimental results, MultiPhaseFormer achieves a mean Dice coefficient of 0.5373 compared to 0.4616 for the baseline 3D U-Net, and a pathology type classification accuracy of 0.80 compared to 0.64. It was found that voxel-wise directed interphase attention reduces inter-class confusion for diagnostically similar tumor types. The obtained results are applicable to clinical decision support systems in radiology and confirm that accounting for the temporal structure of interphase data is a crucial factor in improving the quality of automatic multiphase CT analysis.