Knowledge-based Labeled Data Selection in Semi-Supervised Learning

Motivation: To reach the optimal decision boundary in semi-supervised model learning, labelled data should have all the data diversity representation. Goal(s): To evaluate the impact of knowledge-based labeled data selection in semi-supervised learning of spine vertebrae segmentation model. Approach: We train semi-supervised learning model for using random and knowledge-based labeled data selection method and compared the model performance using Dice score and Hausdorff distance. Results: For metal implant cases, our semi-supervised model trained on knowledge-based labelled data selection, shows an increase of 5% in Dice score and 35% reduction in the Hausdorff distance metric. Impact: We present the importance of data diversity representation in the labelled data of semi-supervised model learning.