Abstract High-frequency oscillations are specific biomarkers of epileptogenic zone, yet their noninvasive localization from scalp electroencephalography (EEG) remains challenging due to their brief duration, low amplitude, and poor signal-to-noise ratio. Here, we introduce TH-DeepSIF, a transformer-based deep learning framework trained on biologically realistic neural mass model simulations for noninvasive high-frequency oscillation source imaging. The model was evaluated in simulated single- and dual-source scenarios, where it recovered both the spatial location and temporal dynamics of high-frequency oscillation generators under increasing spatiotemporal complexity. We further validated TH-DeepSIF in 20 patients with drug-resistant epilepsy by comparing reconstructed high-frequency oscillation sources with surgical resection regions. TH-DeepSIF achieved accurate localization and waveform reconstruction in simulations, demonstrating low localization error and strong temporal correspondence with ground truth. In clinical data, pathological high-frequency oscillations (spike ripples) showed strong agreement with resection regions, with a median localization error of 10.9 mm and specificity of 0.883. In a subset of six patients with intracranial EEG, pathological high-frequency oscillation source estimates were spatially close to the seizure onset zone (median localization error: 2.41 mm). Compared with all detected high-frequency oscillations, imaging based on pathological high-frequency oscillations yielded significantly improved spatial correspondence and performance across multiple metrics. Reconstructed source activity was also temporally aligned with both scalp high-frequency oscillation bursts and associated spike deflections. These results demonstrate that TH-DeepSIF provides a data-driven framework for noninvasive high-frequency oscillation source imaging with improved anatomical specificity and support the use of pathological high-frequency oscillations as informative biomarkers for epileptogenic zone localization.