Related Experiment Videos
Evaluation of transfer learning in deep convolutional neural network models for cardiac short axis slice
1Department of Computer Science and Engineering, Sogang University, Seoul, Republic of Korea.
Scientific Reports
|January 20, 2021
Summary
This study introduces an automated method for classifying cardiac MRI short axis slices using deep convolutional neural networks (CNNs). The fine-tuned VGG16 model demonstrated superior performance, enhancing the accuracy of left ventricle (LV) analysis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Cardiology
Background:
- Accurate segmentation of cardiac MRI data is crucial for quantifying left ventricle (LV) parameters like ejection fraction and mass.
- Standard post-processing guidelines emphasize correct identification of the short axis slice range for reliable quantification.
- Manual slice identification is time-consuming and prone to errors, necessitating automated solutions.
Purpose of the Study:
- To investigate the feasibility of using transfer learning with deep convolutional neural networks (CNNs) for automatic classification of cardiac MRI short axis slice ranges.
- To develop and evaluate an automated system for categorizing slices into 'out-of-apical', 'apical-to-basal', and 'out-of-basal' based on their location within the LV.
- To identify the most effective CNN architecture for this specific cardiac imaging task.
Main Methods:
- Transfer learning was applied to nine popular deep CNN architectures.
- A custom user interface was developed for efficient labeling of cardiac MRI slices to create training datasets.
- The CNNs were trained to classify short axis slices into three predefined categories based on their anatomical location.
- Performance was evaluated using unseen test data.
Main Results:
- The fine-tuned VGG16 model achieved the highest performance across all evaluation metrics.
- The VGG16 model demonstrated robust classification accuracy for cardiac MRI short axis slices.
- The developed method shows promise for automating a critical step in cardiac MRI analysis.
Conclusions:
- Transfer learning with deep CNNs is a feasible and effective approach for automatic cardiac MRI short axis slice range classification.
- The VGG16 architecture, when fine-tuned, is the most suitable model for this task among those evaluated.
- Automating this classification process can improve the efficiency and accuracy of cardiac MRI analysis.
