Functional near-infrared spectroscopy techniques, in the form of either optical topography (OT) or diffuse optical tomography (DOT), can non-invasively recover the hemodynamic changes occurring in the activated cerebral cortex. In comparison with the traditional OT that provides a less quantitative absorption perturbation map along the subject domain surface, a successful DOT has ability to quantify depth-resolved information that relies on abundant boundary overlapping measurements using a high-density (HD) source-detector array. To achieve a trade-off between the temporal resolution and sensitivity by channel cross-talk suppression, a hybrid frequency- and time-division-multiplexing strategy have to be normally adopted to the HD-DOT implementation, where the temporal resolution degradation due to the multi-field illuminations might still prevent from capturing the high frequency information. In this work, a deep-learning based pre-OT method has been proposed to improve the temporal resolution of HD-DOT. The pre-OT could provide prior information on activation regions to exclude measurements of non-sensitive data. We have performed simulation and phantom experiments to evaluate the performances of the proposed method, and demonstrated its superiority over the stand-alone HD-DOT in improving both the temporal resolution and localization accuracy.