Automated Segmentation of Knee Articular Cartilage on MRI Data: Increasing Network Capacity with Transfer Learning

In this study we evaluated the possibility of using transfer learning to improve the segmentation accuracy of femoral and tibial knee articular cartilage of a small locally acquired and annotated dataset. Two conditional Generative Adversarial Networks were trained - one with pretraining on the much larger SKI10 (Segmentation of Knee Images 2010) dataset and the other with random weight initialisation and no pretraining. Pretraining not only increased cartilage segmentation accuracy of the fine-tuned dataset, but also increased the network’s capacity to preserve segmentation capabilities for the pretrained dataset.