Integracja strukturalno-funkcjonalna do lepszego wykrywania subtelnych zmian epileptogennych: ramy wielomodalnego podejścia łączącego EEG i MRI z wiedzą kliniczną

Abstract Purpose: Focal cortical dysplasia (FCD) is a leading cause of treatment-resistant epilepsy and one of the most diagnostically challenging diseases in neuroimaging. Even the best deep learning models available to detect it rely on imaging data alone and achieve only suboptimal accuracy. A key insight motivates the approach taken here: experienced neuroradiologists and epilepsy specialists donot search the entire brain blindly. Instead, they integrate electroencephalography (EEG) and clinical findings that pinpoint the region of abnormal activity where seizures originate, and use them to focus and prioritize the lesion search. This study tests whether an automated pipeline can mirror this search pattern by dynamically adjusting the detection threshold of a 3D nnU-Net according to thelesion location based on different localization scales (hemisphere, lobe, and sub-region). This design measures the upper-bound performance improvement that EEG or clinical priors could offer. Methods: A 3D nnU-Net was trained on 3D T2-FLAIR images from 175 FCD cases. Anatomical search zones at varying scales were built with SynthSeg, an atlas-based parcellation that runs in seconds per case, far below the roughly eight hours of model training per fold. Rather than applying the fixed 0.5 probability threshold everywhere, the pipeline lowered it inside the localization zone to surface faint sub-threshold peaks and raised it to 0.90 outside the zone to suppress false positives, scoring a case as recovered when any of the top five candidate peaks overlapped the lesion. Results: The model detected 86.3% of lesions (mean Dice 0.57), missing 24 cases (13.7%). Restricting a multi-peak search to the lesion lobe recovered 14 of the 24 misses (58%), which a chance baseline confirms is well above the 1.4 expected from a spatially uninformative search (P Conclusion: This study shows that sub-threshold, spatially coherent signal often persists in cases where a standard nnU-Net reports the case as negative. Future work should focus on integrating real EEG and clinical data to improve deep learning detection pipelines.