This paper proposes a decoupled feature query (DFQ) learning scheme for domain generalized medical image segmentation. The goal is to address feature misalignment among cross-domain medical images. To enhance per-channel representation ability and reduce channel redundancy, a relaxed deep whitening transformation (RDWT) is proposed. To learn similar channel-wise feature patterns from different domains, decoupled deep features are used as queries to guide the entire framework. The DFQ scheme is seamlessly integrated into Transformer segmentation models in an end-to-end manner. Extensive experiments show that the proposed method outperforms existing state-of-the-art methods on two standard benchmarks: optic cup/disk segmentation on fundus images and prostate segmentation on MRI. On the Fundus benchmark, the method achieves an average DSC of 89.92% and ASD of 1.07%. On the Prostate benchmark, it achieves an average DSC of 89.92% and ASD of 1.07%. The proposed DFQ scheme significantly improves the performance on both benchmarks, outperforming the second-best methods by up to 1.31% and 1.98% in DSC metrics. The method is effective in reducing channel redundancy and improving feature alignment, leading to better generalization on unseen target domains. The proposed DFQ scheme is integrated into the Transformer segmentation model through a relaxed deep whitening transformation, which helps in learning more expressive and less redundant features. The method is validated through extensive experiments on two standard benchmarks, demonstrating its effectiveness in domain generalized medical image segmentation.This paper proposes a decoupled feature query (DFQ) learning scheme for domain generalized medical image segmentation. The goal is to address feature misalignment among cross-domain medical images. To enhance per-channel representation ability and reduce channel redundancy, a relaxed deep whitening transformation (RDWT) is proposed. To learn similar channel-wise feature patterns from different domains, decoupled deep features are used as queries to guide the entire framework. The DFQ scheme is seamlessly integrated into Transformer segmentation models in an end-to-end manner. Extensive experiments show that the proposed method outperforms existing state-of-the-art methods on two standard benchmarks: optic cup/disk segmentation on fundus images and prostate segmentation on MRI. On the Fundus benchmark, the method achieves an average DSC of 89.92% and ASD of 1.07%. On the Prostate benchmark, it achieves an average DSC of 89.92% and ASD of 1.07%. The proposed DFQ scheme significantly improves the performance on both benchmarks, outperforming the second-best methods by up to 1.31% and 1.98% in DSC metrics. The method is effective in reducing channel redundancy and improving feature alignment, leading to better generalization on unseen target domains. The proposed DFQ scheme is integrated into the Transformer segmentation model through a relaxed deep whitening transformation, which helps in learning more expressive and less redundant features. The method is validated through extensive experiments on two standard benchmarks, demonstrating its effectiveness in domain generalized medical image segmentation.