Accurate segmentation of white matter hypoattenuations on brain CT scans is critical for AI-powered stroke prognosis and research. However, scanner-induced variability significantly limits the generalisability of automated segmentation algorithms across multi-centre datasets. To address this, we propose a simple but robust and interpretable harmonisation approach based on aligning the dominant intensity peak corresponding to normal brain tissue, followed by intensity scaling. We evaluated our method using the nnU-Net framework on 91 hospital-acquired scans from four CT scanner manufacturers. The nnU-Net model trained on images after harmonisation demonstrated statistically significant improvements in segmentation accuracy compared to non-harmonised baselines, notably achieving up to a 56.4% reduction in relative volume difference, 5.15% higher precision, 1.8% improvement in Dice similarity coefficient, and better volumetric agreement, especially in mild-to-moderate lesion cases. Our results highlight the method’s potential to substantially reduce inter-scanner biases, improving segmentation reliability and volumetric consistency, crucial for clinical use and multi-centre research. This harmonisation approach provides a practical preprocessing solution to enhance the robustness of automated segmentation in heterogeneous CT imaging environments.