Parameter-efficient adaptation of foundational models for automated myocardial strain analysis

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated myocardial strain analysis is crucial for diagnosing cardiac conditions.
• Deep learning models often struggle with cross-dataset generalization due to variations in imaging equipment and institutions.
• Efficient adaptation of general-purpose models is needed for reliable cardiac strain estimation.

Purpose of the Study:

• To investigate efficient adaptation of foundational models for cardiac strain estimation.
• To establish minimal data requirements for robust cross-domain deployment of these models.
• To develop a method for accurate segmentation and motion tracking in cardiac images.

Main Methods:

• CardiacSAM, a lightweight adaptation of the Segment Anything Model (SAM) using Low-Rank Adaptation (LoRA), was developed for cardiac structure segmentation.
• UniGradICON was employed for myocardial motion estimation.
• Global strain was computed in cardiac coordinates for both short-axis (SAX) and long-axis (LAX) views.

Main Results:

• CardiacSAM achieved high segmentation accuracy (mean LV Dice score of 0.90 ± 0.02) with minimal training data (10 studies/domain for SAX, 50 for LAX).
• Motion estimation showed a median AEPE of 3.08 mm.
• Strain measurements effectively discriminated between healthy and diseased groups across multiple datasets.
• SAX and LAX strain analysis revealed moderate correlation for radial strain and weak correlation for longitudinal strain, indicating complementary information.

Conclusions:

• Parameter-efficient adaptation of foundational models enables automated cardiac strain assessment.
• The proposed method demonstrates high data efficiency and robust generalization across different datasets.
• This approach facilitates reliable automated cardiac strain analysis in diverse clinical settings.