SAID-Net: enhancing segment anything model with implicit decoding for echocardiography sequences segmentation

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Echocardiography sequence segmentation is crucial for diagnosing cardiac conditions.
• General segmentation models like Segment Anything Model (SAM) struggle with complex cardiac anatomy and ultrasound image subtleties.
• Existing methods lack the adaptability and robustness needed for precise cardiac delineation.

Purpose of the Study:

• To introduce SAID (Segment Anything with Implicit Decoding), a novel framework designed to improve echocardiography segmentation accuracy and robustness.
• To leverage implicit neural representations (INR) integrated with SAM for enhanced cardiac image analysis.
• To address the limitations of current models in handling complex cardiac structures and subtle ultrasound boundaries.

Main Methods:

• Developed SAID, a framework combining SAM with Implicit Neural Representations (INR).
• Employed a Hiera-based encoder for multi-scale feature extraction and a Mask Unit Attention Decoder for detailed cardiac delineation.
• Utilized orthogonalization for feature diversity and I^2 Net for improved handling of misaligned contextual features.

Main Results:

• SAID achieved a Dice Similarity Coefficient (DSC) of 93.2% and Hausdorff Distance (HD95) of 5.02 mm on the CAMUS dataset.
• On the EchoNet-Dynamics dataset, SAID attained a DSC of 92.3% and HD95 of 4.05 mm.
• SAID demonstrated superior performance compared to state-of-the-art methods in echocardiography sequence segmentation.

Conclusions:

• SAID significantly enhances accuracy, adaptability, and robustness in echocardiography segmentation.
• The integration of INR with SAM provides a powerful approach for complex cardiac image analysis.
• SAID represents a significant advancement for automated cardiac image segmentation in clinical cardiology.