# Learning Directional Feature Maps for Cardiac MRI Segmentation

**Authors:** Feng Cheng, Cheng Chen, Yukang Wang, Heshui Shi, Yukun Cao, Dandan Tu, Changzheng Zhang, Yongchao Xu  
**arXiv:** [2007.11349](https://arxiv.org/abs/2007.11349)  
**Date:** 2020-07-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2007.11349v1/

## Key points

- Cardiac MRI segmentation is important for clinical diagnosis of personalized cardiac performance parameters
- Existing methods face challenges in accurately segmenting the heart due to indistinct boundaries and heterogeneous intensity distributions
- Two main issues arise: inter-class indistinction and intra-class inconsistency
- The proposed method leverages directional feature maps to enhance differences between classes and similarities within each class
- The approach involves performing cardiac segmentation and learning a direction field for each pixel, pointing away from the nearest cardiac tissue boundary
- A direction field (DF) module is used to achieve this
- A feature rectification and fusion (FRF) module is introduced to improve original segmentation features and obtain the final result
- These modules are simple yet effective and can be easily integrated into existing segmentation networks without significantly increasing time or computational complexity
- The proposed method is evaluated on two datasets: the 2017 MICCAI Automated Cardiac Diagnosis Challenge (ACDC) dataset and a large-scale self-collected dataset
- Results demonstrate good segmentation performance and robust generalization ability of the proposed method compared to existing methods.

## Summary

Cardiac MRI segmentation is an essential component of clinical diagnosis for evaluating personalized cardiac performance parameters. However, existing methods face challenges in accurately segmenting the heart due to indistinct boundaries and heterogeneous intensity distributions in cardiac MRI images. Specifically, two main issues arise: inter-class indistinction and intra-class inconsistency. To address these challenges, this study proposes a novel method that leverages directional feature maps to enhance the differences between different classes and the similarities within each class. The proposed approach involves performing cardiac segmentation and learning a direction field for each pixel, pointing away from the nearest cardiac tissue boundary. This is achieved through a direction field (DF) module. Based on the learned direction field, a feature rectification and fusion (FRF) module is introduced to improve the original segmentation features and obtain the final segmentation result. Importantly, these modules are simple yet effective and can be easily integrated into existing segmentation networks without significantly increasing time or computational complexity. The proposed method is evaluated on two datasets: the 2017 MICCAI Automated Cardiac Diagnosis Challenge (ACDC) dataset and a large-scale self-collected dataset. The results demonstrate good segmentation performance and robust generalization ability of the proposed method. In conclusion, this study presents a novel approach for cardiac MRI segmentation that addresses the challenges of inter-class indistinction and intra-class inconsistency by exploiting directional feature maps to achieve accurate segmentation results with improved performance compared to existing methods.