We present TorchIO, an open-source Python library for efficient loading, preprocessing, augmentation and patch-based sampling of medical images for deep learning. It follows the design of PyTorch and relies on standard medical image processing libraries such as SimpleITK or NiBabel to efficiently process large 3D images during the training of convolutional neural networks. We provide multiple generic as well as magnetic-resonance-imaging-specific operations for preprocessing and augmentation of medical images. TorchIO is an open-source project with code, comprehensive examples and extensive documentation shared at https://github.com/fepegar/torchio.