CMRCUNet: a cross-modal calibration U-Net for acute ischaemic stroke lesion segmentation from multi-sequence MRI

Abstract Purpose: Acute ischaemic stroke is a major cause of mortality and long-term disability worldwide, and rapid delineation of infarct lesions on brain magnetic resonance imaging (MRI) is important for supporting time-sensitive treatment decisions. Multi-sequence MRI, including diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC) maps, and fluid-attenuated inversion recovery (FLAIR), provides complementary information for characterising acute ischaemic lesions. However, many deep learning approaches process these modalities using shared encoders and direct skip connections, which may limit modality-specific feature learning and remain challenged by the severe foreground–background imbalance characteristic of stroke lesion segmentation. Method: We propose CMRCUNet, a multimodal segmentation framework that combines modality-specific residual encoders, a Cross-Modal Calibration (CMC) module, and an attention-gated decoder. The modality-specific encoders independently learn sequence-dependent representations from DWI, ADC, and FLAIR, while the CMC module performs adaptive, input-dependent weighting of modality-specific features at multiple resolution levels before decoder fusion. Focal Tversky Loss is used to address severe lesion–background imbalance. The framework is evaluated in both 2D and 2.5D configurations; the 2.5D configuration incorporates neighbouring axial slices to provide limited through-plane contextual information while retaining two-dimensional convolutional operations. Test-time augmentation (TTA) is additionally evaluated to improve inference robustness. Results: CMRCUNet was evaluated on the ISLES 2022 dataset using a patient-level evaluation protocol with an independent 37-patient test cohort. The 2.5D configuration combined with TTA achieved a Dice similarity coefficient of 0.792 ± 0.142, sensitivity of 0.854 ± 0.132, specificity of 0.996 ± 0.012, and a 95th-percentile Hausdorff distance (HD95) of 8.21 ± 8.70 mm. Conclusion: CMRCUNet provides a multimodal framework that integrates modality-specific feature extraction, adaptive cross-modal calibration, attention-gated decoding, and limited inter-slice contextual information for acute ischaemic stroke lesion segmentation. The results support further investigation of adaptive cross-modal calibration, particularly through external multi-centre validation.