Generalizability of nnU-Net for automatic segmentation of knee MRI

To investigate the generalizability of deep learning segmentation models, three different nnU-Nets (2D, 3D, and ensemble) were trained on the OAI dataset and tested on a different dataset. In addition to the nnU-Nets trained on the original OAI data, the style of the test set was transferred to the training set using a CycleGAN method and the nnU-Nets were trained again. Depending on the tissue, the 3D nnU-Net or the ensemble trained on the original or stylized training data had the highest segmentation accuracy in the test set. The results indicate that nnU-Net may generalize well to independent datasets.