Rapid diagnosis is crucial to cancer treatment. Multi-core fibers (MCF) are slim and enable access to deep tissue for all-optical biopsies in real-time. Deep neural network (DNN)-based image reconstruction can depixelate structural artifacts and enhance spatial resolution beyond physical limitations. However, limited available tissue samples for training and simulator performance for MCF imaging make DNN-assisted fiber endoscopy for in vivo brain cancer diagnosis still challenging. Here we propose using a digital twin to solve these difficulties. A dataset is simulated for pre-training a MCF reconstruction network. Then transfer learning is applied to the network on a few ten experimental image pairs of glioblastoma samples. The results show, even when transfer learning on only 10 experimental image pairs, the mean structural similarity value of the test set can be increased from 0.547 to 0.932. The reconstructed images can achieve a comparable resolution with widefield microscopes when increasing the data size to 50, which are hundreds of times less than the training from scratch. The digital twin provides an efficient and highly feasible approach towards minimally invasive all-optical biopsies in neurosurgeries.