A unified frequency-domain framework for tilted slice localization and ischemic stroke detection

Biomedical Signal Processing and Control. 2027. Vol. 129. P. 111284–111284.

Khodadoust J., Kulikova S., Khodadoust F.

Acute ischemic stroke (AIS) analysis from two-dimensional (2D) clinical imaging is hindered by uncontrolled slice tilt and geometric inconsistencies that violate the assumptions of pose-agnostic deep learning (DL) models. This paper proposes a unified geometry-aware, frequency-domain framework for tilted slice localization and ischemic stroke segmentation that explicitly decouples pose estimation from lesion analysis. The method first recovers the full six-degree-of-freedom (6-DoF) rigid pose of arbitrarily tilted slices using frequency-domain slice-to-volume registration, converting a geometrically ill-posed segmentation problem into an anatomically normalized one. Building on this normalization, a frequency-domain segmentation network is introduced that exploits Hermitian symmetry and discriminative spectral bands to enhance sensitivity to ischemic tissue, complemented by a teacher–student knowledge distillation (KD) strategy to improve generalization. Extensive experiments on four public benchmark datasets across multiple imaging modalities demonstrate consistent state-of-the-art (SOTA) performance under controlled geometric variability, with notable improvements on the clinically critical core–penumbra segmentation task and preliminary evidence of robustness to certain domain shifts. The results confirm that explicit geometric modeling combined with spectral-domain analysis provides a robust foundation for medical image segmentation under controlled geometric variability.

Keywords: нейронные сети, MRI, MNI normalization, пространственная нормализация, neuronal networks, МРТ