Unbalanced graph-based transduction on superpixels for automatic cervigram image segmentation

We propose a novel medical image segmentation algorithm by transductively inferring the labels. In this approach, superpixels are first generated to incorporate the local spatial information and also to speed up the segmentation. The segmentation task can be deemed as an unbalanced superpixels labeling problem due to the fact that the region of interest is only a small fraction compared to the whole image. We present a new transductive learning-based algorithm called Class Averaging Graph-based Transduction (CAGT) to avoid the biased labeling caused by the imbalance. The proposed algorithm was applied to the automatic cervigram image segmentation to demonstrate it effectiveness.