RAMAS-Net: a module-optimized convolutional network model for aortic valve stenosis recognition in echocardiography.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Aortic stenosis (AS) is a critical valvular heart disease impairing cardiac function.
• Echocardiography is vital for AS diagnosis but faces accuracy challenges due to variability and image quality.
• Accurate AS diagnosis and severity classification are crucial for effective patient management.

Purpose of the Study:

• To develop and evaluate a deep learning model for automated AS identification and diagnosis using echocardiography.
• To enhance diagnostic accuracy and efficiency in assessing aortic stenosis.
• To provide a reliable tool for healthcare professionals to support clinical decision-making.

Main Methods:

• A deep learning model, RSMAS-Net, was proposed, enhancing the ResNet50 architecture.
• The model incorporated Spatial and Channel Reconstruction Convolution (SCConv) and Multi-Dconv Head Transposed Attention (MDTA) modules.
• RSMAS-Net was trained and validated on the TMED-2 echocardiography dataset.

Main Results:

• RSMAS-Net achieved 94.67% accuracy and an AUC of 0.95 for AS identification on TMED-2.
• The model demonstrated strong performance in AS severity classification with an AUC of 0.93.
• RSMAS-Net outperformed baseline models in key metrics including recall, precision, and inference time, also showing good results on TMED-1 (AUC 0.91).

Conclusions:

• RSMAS-Net effectively diagnoses and classifies AS severity from echocardiographic images.
• The SCConv and MDTA modules improve diagnostic accuracy and reduce model complexity.
• The model shows significant potential for enhancing AS assessment and clinical decision support.