Semi-supervised Learning in Medical Image Analysis

As a branch of machine learning, Convolutional Neural Networks (CNN) have shown great potential in the medical image analysis field in recent years. CNN can extract rich 2D or higher-dimensional information from the original medical images and derive numerous image features at multiple scales automatically. With a well-designed training strategy, CNN has achieved state-of-the-art segmentation results in very short computation times in many applications. However, a problem of CNN is that in most cases, successful model training requires a large annotated dataset. Such datasets are difficult and costly to obtain. An alternative could be to apply semi-supervised learning (SSL), in which the model is learned from a combination of annotated and unannotated data. This has the advantage that unannotated data is usually much easier to obtain in practice. Using semi-supervised learning, fewer manually annotated images may be required for successful application of CNN.

We will, therefore, investigate the possibility of combining CNN and semi-supervised learning and see if we could make improvements to the algorithm on different tasks. The method has been applied and get state-of-the-art results in three different segmentation tasks: aorta and pulmonary artery segmentation, white matter hyperintensities segmentation and ischemic stroke lesion segmentation. The semi-supervised learning research will keep focusing on the same applications in order to make a fair comparison with the current supervised method as well as the previous SSL approaches. The future work will be proposing a new efficient SSL method that could take advantage of autoencoder, CNN and self-supervision. The research could benefit clinical applications which have a large amount of unlabeled data.

Research line

Model-based Medical Image Analysis

Associated people

Faculty

PhD students