nnU-Net for the Automatic Knee Segmentation from CT Images: A Comparative Study with a Conventional U-Net Model

This study aims at comparing the nnU-Net, an open-source deep learning framework, with a previous customized U-Net model that we developed for the automatic segmentation of tibial and femoral bones from CT scans. The main purpose of our work is to develop a segmentation module that could be integrated into a surgical planning software for the design of customized Total Knee Prosthesis. The nnU-Net framework was chosen for its user-friendly design and features developed for medical imaging. The same dataset of 112 CT scans of lower limbs from 63 patients was used to train and test both our customized U-Net model and the nnU-Net model. All these data were manually annotated. The evaluation was done by computing the Average Symetric Surface Distance, the Dice Coefficient, the Hausdorff Distance, the precision, the recall and the Jaccard Index. Both models yielded similar results on these metrics, but the nnU-Net model is easier to setup. The performances of both models are also consistent with the literature, however, further tests on pathological data will be needed.