A method for dose-gradient based optimization of an intensity modulated radiation therapy plan. First, an optimizer (6) performs a first optimization (40) of the plan to generate dose distributions corresponding to the plan. Next, the optimizer (6) generates a beam specific dose gradient map (42) for each beam of the plan. Then, new dose gradients are specified (44) for the plan. Last, the optimizer (6) performs a final optimization (46) using the new dose gradients. The final optimization is given the new dose gradients as soft constraints into an objective function. The optimizer (3) applies a limiting factor to the objective function such that a first dose gradient is limited by the optimizer only if the first dose gradient exceeds the new dose gradient for a specific beamlet.