Personalized echocardiographic segmentation via bidirectional encoder representations from transformers Y-shaped...

Area of Science:

• Medical imaging
• Artificial intelligence in cardiology
• Quantitative echocardiography

Background:

• Accurate cardiac segmentation is crucial for assessing heart function but is limited by manual annotation subjectivity and automated methods' lack of generalizability.
• Existing automated segmentation tools often fail to account for inter-patient anatomical variations, such as sex-related differences, hindering their clinical application.
• Personalized and anatomically adaptive segmentation is needed to overcome limitations in current echocardiographic analysis.

Purpose of the Study:

• To introduce BTY-Net, a novel framework for automatic echocardiogram segmentation that incorporates patient-specific attributes.
• To enable personalized and anatomically adaptive cardiac structure segmentation using a Bidirectional Encoder Representations from Transformers (BERT) Text-based Y-shaped Network (BTY-Net).
• To improve the generalizability and accuracy of automated echocardiographic segmentation by addressing inter-patient variability.

Main Methods:

• BTY-Net utilizes a Unet3+ backbone with a Transformer encoder and a multi-layer denoising filter.
• A pre-trained BERT model encodes patient demographic and acquisition context as natural language embeddings for personalized segmentation.
• The framework was evaluated on the Cardiac Acquisitions for Multi-structure Ultrasound segmentation dataset, benchmarking against eight state-of-the-art models using Dice similarity and Hausdorff Distance.

Main Results:

• BTY-Net achieved superior segmentation performance with the highest Dice coefficients (e.g., LV endocardium: 0.9316) and lowest Hausdorff Distances across key cardiac structures.
• The model demonstrated significant improvements over the strongest baseline, enhancing Dice by up to 0.02-0.03 and reducing HD by 1.1-1.5 mm.
• BTY-Net showed high agreement with reference ejection fraction (correlation=0.9119) and maintained stable performance across male and female subgroups, confirming its robustness to anatomical diversity.

Conclusions:

• BTY-Net provides an effective, interpretable, and personalized solution for echocardiographic analysis.
• The multimodal fusion of patient information and image data enhances segmentation accuracy and provides clinically meaningful attention maps.
• BTY-Net offers a sex-robust, clinically interpretable framework for routine echocardiographic analysis.