Cardiac Disease Classification Using Two-Dimensional Thickness and Few-Shot Learning Based on Magnetic Resonance...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiac cine MRI is vital for noninvasive cardiac function assessment.
• Deep learning faces challenges in cine MRI analysis due to limited data.
• Classical machine learning often requires handcrafted features and clinical data.

Purpose of the Study:

• To develop a novel method for classifying heart disease, specifically cardiomyopathy, using only segmented cardiac cine MRI output maps.
• To overcome data limitations in deep learning for cardiac MRI analysis.
• To enable rapid analysis and diagnosis of heart conditions.

Main Methods:

• Modified a fully convolutional EfficientNetB5-UNet for semantic segmentation of MRI slices.
• Employed a 2D thickness algorithm to create end-diastole (ED) and end-systole (ES) cardiac volume representations.
• Utilized a few-shot model with an adaptive subspace classifier for classification based on thickness images.

Main Results:

• Achieved high segmentation performance: average Dice coefficients of 96.24% (ED) and 89.92% (ES) for the left ventricle (LV).
• Reported Dice coefficients for the right ventricle (RV) as 92.90% (ED) and 86.92% (ES), and for myocardium as 88.90% (ED) and 90.48% (ES).
• Attained 92% accuracy in classifying cardiomyopathy groups without relying on clinical features.

Conclusions:

• A novel, rapid analysis approach for heart disease diagnosis, particularly for cardiomyopathy, was successfully developed using segmented cardiac cine MRI output maps.
• The proposed method demonstrates high accuracy and segmentation performance, overcoming data limitations.
• This technique offers a promising alternative for efficient and accurate cardiac condition diagnosis.