Unsupervised Representation Learning of Medical Images for Downstream Segmentation

Mingrui Zhuang, Hongkai Wang

Keywords: Medical image segmentation; Representation learning; Incomplete supervision learning

TL;DR: Training an unsupervised anatomical position encoding network to guide the downstream segmentation network.

Abstract: Automatic abdominal organ and tumor segmentation from CT scans can enhance clinical diagnosis and treatment planning, but manual annotation remains predominant due to limitations in current automated methods' robustness and accuracy. We propose a novel representation learning approach that trains a large-scale anatomical positional encoding network (LAPN) in an unsupervised manner, overcoming limited labeled data. LAPN encodes pixel-level anatomical localization information to guide downstream segmentation. We employed a U-Net network that takes in both the original image and positional encoding features to accomplish the specific segmentation task. Due to the large size of the LAPN segmentation pipeline, we use knowledge distillation to train a lightweight U-Net for efficient inference. The experiments demonstrate that LAPN can leverage unlabeled data to improve the performance of the segmentation network, particularly for organs with relatively fixed anatomical positions.

Submission Number: 33