Lens-based mmWave massive MIMO can significantly reduce the number of required RF chains without obvious performance loss, where the accurate information of beamspace channel is required. However, existing beamspace channel estimation schemes are based on the 2D beamspace channel model. In this paper, we consider the more general 3D beamspace channel model, and propose an adaptive support detection (ASD)-based channel estimation scheme. The basic idea is to decompose the 3D beamspace channel estimation problem into several sub-problems, and each one only deals with a sparse channel component. For each channel component, we first adaptively detect its support with high accuracy by exploiting the horizontal and vertical sparsity of 3D beamspace channel. Then, we remove the influence of this channel component to detect the support of the next channel component. After the support detections of all channel components, we can estimate the nonzero elements of the beamspace channel with low pilot overhead. Simulation results verify that the proposed scheme enjoys satisfying accuracy, even with low SNR.