Functional near-infrared spectroscopy (fNIRS) has gained interest in the development of brain-computer interface (BCI) for working memory (WM) training. Amplitude averaging of oxygenated hemoglobin (oxy-Hb) signal over the predefined region of interest (ROI) is typically used to compute WM load. It is unclear to what extent extracerebral artifacts can affect WM assessment. To examine this, a technique adopting multi-distance probe configuration and independent component analysis (MD-ICA) was applied to split the original fNIRS signals into hemodynamic signals originating from the deep (cerebral) and shallow (extracerebral) tissue layers. Twenty-five healthy participants performed letter 1- and 2-back tasks, symbolizing low and high WM load respectively. In the bilateral dorsolateral prefrontal cortex (DLPFC), increasing WM load evoked significant changes in both original and deep oxy-Hb activation, but not in the shallow oxy-Hb. Under low WM load, the bilateral DLFPC activation was significantly higher than that of the middle prefrontal cortex (mPFC), only seen through the deep signal. Conversely, under high WM load, the significant difference in brain activation between the bilateral DLPFC and mPFC were observed in both original and deep oxy-Hb. This highlights extracerebral artifacts correction might be useful when searching for activation regions. However, when activation areas are known, the signal intensities in original oxy-Hb are sufficiently profound to quantify and differentiate WM load. © 2020 IEEE.