SymBreak: Breaking the Implanted False Belief to Identify Subvisual Ischemic Changes

Stroke is a leading cause of death and disability, and rapid diagnosis is critical to improving patient outcomes. Although non-contrast CT (NCCT) is the first-line imaging modality for suspected stroke, hyperacute ischemic changes, particularly within 4.5 hours, remain difficult to detect because of their subtle imaging manifestations. When pathological signals are extremely weak, lesion-related responses can be difficult to distinguish from the model's own background hallucinations—weak, spatially irregular lesion-like responses elicited even by negative scans. Here, we propose SymBreak, a new training and inference paradigm based on false-belief implantation that organizes these otherwise unstructured background responses into a predictable pattern, thereby improving the detectability of subtle ischemic changes. During training, generative models transform unilateral ischemic samples into bilaterally symmetric counterfactual cases, implanting the false belief that ischemic changes MUST BE symmetric between hemispheres. During inference, background responses in non-stroke cases tend to follow this implanted symmetry, whereas genuine unilateral ischemic changes break it. We quantify this violation using the Symmetry-Breaking Index (SBI), which converts hemispheric imbalance in predicted lesion evidence into a patient-level diagnostic signal. Across five public NCCT datasets, SymBreak consistently improves patient-level discrimination over existing intensity- and lesion-size-based methods across four held-out AIS cohorts and multiple segmentation architectures. Response analyses further show that false-belief implantation substantially reduces hemispheric imbalance in non-stroke cases while preserving pronounced symmetry breaking in unilateral AIS. Together, these findings establish SymBreak as an accurate and robust approach for detecting hyperacute ischemic changes and suggest a broader weak-signal learning paradigm: deliberately structuring model background responses enables pathology to be detected through violations of the resulting structure.