Attention-UNet architectures with pretrained backbones for multi-class cardiac MR image segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning significantly improves medical image segmentation, aiding diagnosis and treatment by automating tasks like Computed Tomography (CT) and Magnetic Resonance Imaging (MRI).
• Cardiac Magnetic Resonance Imaging (CMRI) segmentation is crucial for heart measurements, but deep learning models face challenges like pixel homogeneity in adjacent organs.
• Attention mechanisms offer a promising approach to enhance automated segmentation algorithms in medical imaging.

Purpose of the Study:

• To investigate the impact of attention mechanisms on the UNet model for cardiac image segmentation.
• To compare the performance of Attention-UNet with and without pretrained backbone networks (ResNet50, DenseNet121).

Main Methods:

• Implemented and evaluated three UNet-based networks: Attention-UNet, Attention-UNet with ResNet50 backbone, and Attention-UNet with DenseNet121 backbone.
• Utilized the ACDC Challenge 2017 dataset for experiments.
• Performance was assessed using Dice Coefficient, Intersection over Union (IoU) Coefficient, and cross-entropy loss.

Main Results:

• Attention-UNet achieved a Dice Coefficient of 0.9889 and IoU of 0.9781.
• Attention-UNet with DenseNet121 backbone yielded a Dice Coefficient of 0.9801 and IoU of 0.9612.
• Attention-UNet with ResNet50 backbone resulted in a Dice Coefficient of 0.9720 and IoU of 0.9457.

Conclusions:

• The attention mechanism significantly enhances UNet performance in CMRI segmentation.
• Pretrained backbone networks offer varying degrees of improvement, with DenseNet121 showing strong results.
• The proposed methods are competitive with, and potentially superior to, state-of-the-art techniques for cardiac image segmentation.