ScribSD: Scribble-Supervised Fetal MRI Segmentation Based on Simultaneous Feature and Prediction Self-distillation

Abstract

Automatic segmentation of different organs in fetal Magnetic Resonance Imaging (MRI) plays an important role in measuring the development of the fetus. However, obtaining a large amount of high-quality manually annotated fetal MRI is time-consuming and requires specialized knowledge, which hinders the widespread application that relies on such data to train a model with good segmentation performance. Using weak annotations such as scribbles can substantially reduce the annotation cost, but often leads to poor segmentation performance due to insufficient supervision. In this work, we propose a Scribble-supervised Self-Distillation (ScribSD) method to alleviate this problem. For a student network supervised by scribbles and a teacher based on Exponential Moving Average (EMA), we first introduce prediction-level Knowledge Distillation (KD) that leverages soft predictions of the teacher network to supervise the student, and then propose feature-level KD that encourages the similarity of features between the teacher and student at multiple scales, which efficiently improves the segmentation performance of the student network. Experimental results demonstrate that our KD modules substantially improve the performance of the student, and our method outperforms five state-of-the-art scribble-supervised learning methods.