High-density (HD) diffuse optical tomography (DOT), as an advanced modality of functional near-infrared spectroscopy, is finding increasing applications in neuroimaging regime. However, it is a primary challenge that the superficial physiological interferences usually significantly contaminate the functional activation reconstruction. In addition, the random noises, majorly the photon-shot and instrumental ones, also cast negative influences on the measurements, and further distort the reconstructed image. To mitigate the adversities, we herein propose a combined scheme of two-layer semi-three-dimensional (S3D) reconstruction and multi-wavelength image fusion, which leverages a mathematical model with explicit physical significance, to suppress the physiological interferences and random noises in HD-DOT reconstruction, respectively. The approach is purely data-driven without additional auxiliary measurement, and comprised of two steps: First, the absorption perturbations are topographically reconstructed over both the scalp (SC) and cerebral-cortex (CC) layers using the two-layer S3D scheme, of which the superficial interferences are estimated from the SC reconstruction and adaptively filtered out from the CC one; Second, the interference-suppressed multi-wavelength CC-images are decomposed using the discrete wavelet transform, and fused at multi-resolutions into a mask for further removal of the random noises. We comprehensively validate the proposed scheme using simulations and phantom experiments, and demonstrate its sound effectiveness in suppressing the physiological interferences and random noises. The performance improvement rather than by more cycles or longer sampling time offers additional payoff: shorter measurement time or higher temporal resolution.