# Self-Supervised Learning for Cardiac MR Image Segmentation by Anatomical Position Prediction

**Authors:** Wenjia Bai, Chen Chen, Giacomo Tarroni, Jinming Duan, Florian Guitton, Steffen E. Petersen, Yike Guo, Paul M. Matthews, Daniel Rueckert  
**arXiv:** [1907.02757](https://arxiv.org/abs/1907.02757)  
**Date:** 2019-07-05  
**Summary page:** https://www.summarizepaper.com/en/paper/1907.02757v1/

## Key points

- CNNs have revolutionized medical image analysis
- Acquiring annotated data is expensive and limited by availability of experts
- Researchers have focused on unsupervised, weakly-supervised, and self-supervised feature learning techniques
- Proposed method for training a cardiac MR image segmentation network using self-supervised learning
- Self-supervised learning achieves high segmentation accuracy comparable to or better than U-net trained from scratch, especially with limited data
- Proposed method improves mean Dice metric from 0.811 to 0.852 for short-axis image segmentation compared to baseline U-net model with only five annotated subjects available
- Self-supervised learning generally produces better segmentations compared to U-net trained from scratch when limited training data is available
- Study highlights potential of self-supervised learning in improving cardiac MR image segmentation

## Summary

Convolutional neural networks (CNNs) have revolutionized medical image analysis by learning discriminative features for classification and regression tasks. However, acquiring large amounts of annotated data is expensive and limited by the availability of expert analysts. To address this challenge, researchers have focused on unsupervised, weakly-supervised, and self-supervised feature learning techniques that utilize abundant unlabeled data without manual annotation. In this study, a novel method for training a cardiac MR image segmentation network is proposed. The network learns features in a self-supervised manner by predicting anatomical positions, which serve as a supervisory signal without requiring additional manual annotation. The results demonstrate that this approach provides a strong signal for feature learning, achieving high segmentation accuracy comparable to or better than a U-net trained from scratch, especially with limited data. For instance, when only five annotated subjects are available, the proposed method improves the mean Dice metric from 0.811 to 0.852 for short-axis image segmentation compared to the baseline U-net model. Additionally, experiments on long-axis image segmentation show that self-supervised learning generally produces better segmentations compared to U-net trained from scratch when limited training data is available. Overall, this study highlights the potential of self-supervised learning in improving cardiac MR image segmentation and suggests exploring other anatomically meaningful pretext tasks to enhance data efficiency in medical imaging applications.