Abstract <p>Time since stroke onset (TSS) is an important basis for treatment decision-making in acute ischemic stroke (AIS). In clinical practice, the DWI–FLAIR mismatch method is used to estimate TSS by assessing differences between stroke-related regions on DWI and FLAIR images. However, substantial modality-style differences between DWI and FLAIR, together with the difficulty of accurately aligning key lesion features, make it challenging for existing methods to quantify this mismatch precisely. To address this issue, we propose StrokeDiffNet. It maps DWI and FLAIR features into a common feature space (CFS) to enable quantitative measurement of inter-modality differences. It also reduces modality-style interference and improves key feature alignment, thereby supporting TSS classification based on difference features. For DWI–FLAIR feature mapping, the DWI and FLAIR encoders are trained with self-reconstruction and interactive supervision. This promotes feature alignment while preserving reconstruction capability and modality-specific characteristics. We further introduce a cross-domain mixed-noise strategy for self-supervised training. This strategy reduces encoder sensitivity to modality-style differences and alleviates their interference with feature alignment. In addition, a key-feature alignment strategy is proposed. It identifies representative regions through correlation and importance analysis of encoder-extracted feature vectors, followed by enhanced alignment training on these key-region features. Experimental results show that StrokeDiffNet achieves strong TSS classification performance, with accuracy, precision, F1-score, and AUC of 73.0%, 78.3%, 78.3%, and 71.3%, respectively, outperforming other mainstream classification networks.</p> Graphical abstract <p></p>