nn-TransUNet: An Automatic Deep Learning Pipeline for Heart MRI Segmentation

Area of Science:

• Medical imaging
• Artificial intelligence
• Cardiovascular medicine

Background:

• Cardiovascular disease (CVD) poses a significant mortality risk.
• Accurate segmentation of cardiac magnetic resonance imaging (MRI) is crucial for CVD diagnosis.
• Current deep learning methods for medical image segmentation are often manual, time-consuming, and dataset-specific.

Purpose of the Study:

• To propose nn-TransUNet, an automated deep learning pipeline for cardiac MRI segmentation.
• To combine the automated experiment planning of nnU-Net with the TransUNet architecture for improved segmentation performance.
• To reduce the manual effort and time required for parameter and hyperparameter tuning in medical image segmentation tasks.

Main Methods:

• nn-TransUNet utilizes a hybrid encoder with vision transformers and convolution layers, and a decoder composed of convolution layers.
• An automatic experiment planning pipeline generates adaptive preprocessing and network training strategies.
• The pipeline integrates nnU-Net's automated planning with TransUNet's network architecture for heart MRI segmentation.

Main Results:

• nn-TransUNet achieved state-of-the-art performance in heart MRI segmentation on the Automatic Cardiac Diagnosis Challenge (ACDC) Dataset.
• The automated pipeline significantly reduced the time and effort associated with manual parameter and hyperparameter optimization.
• The proposed method demonstrates high performance across different datasets, overcoming limitations of single-task specific architectures.

Conclusions:

• nn-TransUNet offers an efficient and effective automated solution for cardiac MRI segmentation.
• The integration of automated experiment planning with advanced network architectures streamlines the research process.
• This approach has the potential to alleviate the workload for researchers and improve diagnostic accuracy for cardiovascular diseases.