Multitask Learning for Estimating Multitype Cardiac Indices in MRI and CT Based on Adversarial Reverse Mapping

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Estimating multitype cardiac indices from cardiac MRI and CT is crucial for comprehensive cardiac function assessment.
• Current models are limited to a single imaging modality (MRI or CT) and lack cross-transferable capabilities.
• Developing a unified approach for cardiac index estimation across different imaging modalities is a significant clinical need.

Purpose of the Study:

• To propose a novel multitask learning method with reverse inferring for estimating multitype cardiac indices from both MRI and CT images.
• To enable parameter transfer from MRI to CT, overcoming the limitations of single-modality models.
• To enhance the clinical utility of cardiac imaging for function assessment through a versatile estimation framework.

Main Methods:

• A multitask learning framework with a reverse mapping network was developed, mapping cardiac indices to images.
• Adversarial training was employed to learn and share task dependencies among multitask learning networks.
• Parameters learned from cardiac MRI were transferred to cardiac CT for cross-modality application.

Main Results:

• The proposed framework demonstrated excellent performance in estimating multitype cardiac indices across both MRI and CT modalities.
• Ten-fold cross-validation on over 2900 cardiac MRI images optimized the framework's performance.
• Independent testing on 2360 cardiac CT images validated the model's effectiveness and transferability.

Conclusions:

• The adversarial reverse mapping multitask learning method effectively estimates multitype cardiac indices from both MRI and CT.
• The study successfully demonstrated cross-modality transfer learning from MRI to CT for cardiac index estimation.
• This approach offers a promising avenue for comprehensive and versatile cardiac function assessment using medical imaging.