Joint Automated analysis of multiphase CT scans, acquired at different time points after contrast agent administration, is highly important for accurate pathology detection. However, it faces a fundamental problem of spatial misalignment between phases due to patient breathing and movement, which leads to a significant reduction in the accuracy of automatic segmentation. Existing pre-registration (phase alignment) methods require manual parameter tuning and are not integrated into trainable pipelines, hindering the full automation of the segmentation process. The aim of this study was to develop a differentiable method for aligning multiphase CT images based on trainable linear affine transformations, fully embedded into the segmentation model training pipeline. Unlike traditional approaches, the proposed method implements a differentiable affine registration module (including translation, rotation, and scaling), whose parameters are optimized via gradient descent without manual tuning and are directly integrated into the computational graph of the nnU-Net model. The study compared the proposed method with a baseline multiphase segmentation approach (based on simple phase concatenation without registration) and an affine registration method implemented using the ITK.Elastix library on the open abdominal CT dataset WAW-TACE. Compared to segmentation without registration, a substantial improvement in quality metrics was achieved: an increase in the Dice coefficient by 43.19% and in the ROC-AUC metric by 13.55%. Compared to Elastix, the improvements were 13.43% in Dice, 6.86% in ROC-AUC, and 19.6% in the accuracy of pathology count detection per CT scan. The practical significance of the research lies in the development of a ready-to-use PyTorch module for integration into CT image analysis pipelines. It enables fully automated registration without the need for manual hyperparameter tuning and offers high computational efficiency
