Left Ventricle Segmentation and Quantification from Cardiac Cine MR Images via Multi-task Learning.

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Accurate left ventricle segmentation and cardiac function quantification are vital for diagnosing cardiovascular diseases.
• Current methods often address segmentation and quantification independently, potentially limiting performance.
• A unified approach could enhance data representation and model generalization.

Purpose of the Study:

• To develop and evaluate a multi-task learning framework using convolutional neural networks (CNNs) for simultaneous segmentation and functional analysis of the left ventricle.
• To leverage probabilistic formulation for learning task uncertainties and adaptively weighting tasks during training.
• To improve the accuracy and generalization of cardiac image analysis.

Main Methods:

• A novel multi-task learning approach employing CNNs was developed.
• The network performs simultaneous segmentation of the left ventricle myocardium and quantification of cardiac functions.
• A probabilistic formulation was used to learn task uncertainties and automatically compute task weights.
• Five-fold cross-validation was conducted on 97 patient 4D cardiac cine-MRI datasets from the STA-COM LV segmentation challenge.

Main Results:

• The proposed multi-task network achieved a Dice overlap of 0.849 ± 0.036 for myocardium segmentation.
• A mean surface distance of 0.274 ± 0.083 mm was obtained for segmentation accuracy.
• Simultaneous estimation of myocardial area yielded a mean absolute difference error of 205 ± 198 mm².

Conclusions:

• The multi-task learning approach effectively performs simultaneous left ventricle segmentation and cardiac function estimation.
• This integrated method demonstrates improved performance and generalization compared to independent task approaches.
• The findings suggest a promising direction for enhancing cardiovascular disease diagnosis through advanced AI in medical imaging.