We investigate the impact of source-side 3D velocity structure on teleseismic travel-time in back projection (BP) analysis of large earthquakes. We use travel-time data of teleseismic events recorded by the Hi-Net array to reveal how travel-time errors vary with source location. In a source area of a few hundred km, where travel-time error varies dominantly linearly, we propose a new interpolation scheme using earthquakes located around the rupture of the main-shock to calibrate the travel-time error, and validate it by relocating inland M>5.0 earthquakes in central Japan. We then apply it to image the rupture of the 2002 Denali earthquake. The calibrated BP result shows that most of the high-frequency radiators are <15 km away from the surface rupture trace. The new result reveals that the rupture started on the Susitna Glacier Fault with a speed of ~1.4km/s, then propagated onto the Denali Fault and accelerated to a super-shear speed approaching the crustal P-wave velocity at approximately 30km. The location of super-shear transition and rupture speed in BP are highly consistent with that inferred from the timing and amplitude ratio of the super-shear and trailing Rayleigh pulses observed on the near fault PS-10 station. Subsequently, the rupture stagnated for ~15s before penetrating through the largest asperity, re-accelerated to a speed of ~5.2 km/s and continued on the last 60 km of the Denali fault and part of Totschunda Fault. This application shows the great potential of the new BP calibration strategy to refine the rupture imaging of other mega-earthquakes.