The terahertz (THz) band radio access with larger available bandwidth is anticipated to provide higher capacities for next-generation wireless communication systems. However, higher path loss at THz frequencies significantly limits the wireless communication range. Massive multiple-input multiple-output (mMIMO) is an attractive technology to increase the Rayleigh distance by generating higher gain beams using low wavelength and highly directive antenna array aperture. In addition, both far-field and near-field components of the antenna system should be considered for modelling THz electromagnetic propagation, where the channel estimation for this environment becomes a challenging task. This paper proposes a novel channel estimation method using a recursive information distillation network (RIDNet) together with orthogonal matching pursuit (OMP) for hybrid-field THz mMIMO channels, including both far-field and near-field components. The simulation experiments are performed using the ray-tracing tool. The results indicate that the proposed RIDNet-based method consistently provides lower channel estimation errors compared to the conventional OMP algorithm for all signal-to-noise ratio (SNR) regimes, and the performance gap becomes higher at low SNR regimes. Furthermore, the results imply that the same error performance of the OMP can be achieved by the RIDNet-based method using a lower number of RF chains and pilot symbols.