Liver tumor segmentation based on semi-supervised learning

Liver tumor segmentation is critical in the diagnosis of liver-related diseases, surgical planning, and prognosis assessment. Nonetheless, acquiring extensive datasets with precise annotations remains a significant challenge. Semi-supervised learning (SSL) is promising for this issue. This paper introduces an SSL technique that leverages feature perturbation to incorporate unlabeled data by applying perturbations at both the image and feature levels. Experimental results reveal that this framework attains a Dice coefficient of 81.2% and an intersection over union of 82.3%, outperforming several traditional SSL techniques and fully supervised training models.