| Čoahkkáigeassu: | With the rise of deep learning applications to medical imaging, there has been a growing appetite for large and well-annotated datasets, yet annotation is time-consuming and hard to come by. In this work, we train a 3D semantic segmentation model in an advanced semi-supervised learning fashion. The proposed SSL framework consists of three models: a Student model that learns from annotated data and a large amount of raw data, a Teacher model with the same architecture as the student, updated by self-ensembling and which supervises the student through pseudo-labels, and an Examiner model that assesses the quality of the student’s inferences. All three models are built with 3D convolutional operations. The overall framework mimics a collaboration between a consistency training Student ↔ Teacher module and an adversarial training Examiner ↔ Student module. The proposed method is validated with various evaluation metrics on a public benchmarking 3D MRI brain tumor segmentation dataset. The experimental results of the proposed method outperform pre-existing semi-supervised methods. The source code, baseline methods, and dataset are available at https://github.com/ziyangwang007/CV-SSL-MIS.
|
|---|