Bi-DCNet: Bilateral Network with Dilated Convolutions for Left Ventricle Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Left ventricular segmentation is crucial for assessing cardiac function using echocardiography.
• Manual segmentation is time-consuming and prone to observer bias.
• Deep learning offers automated segmentation but often overlooks semantic information.

Purpose of the Study:

• To develop an automated deep learning model for accurate left ventricular segmentation in echocardiography.
• To address the limitations of existing deep learning methods by incorporating semantic information.
• To evaluate the proposed model on a large clinical dataset.

Main Methods:

• A novel deep neural network architecture, Bi-DCNet, based on BiSeNet was proposed.
• The model features a spatial path for low-level features and a context path for high-level semantic features.
• Dilated convolutions were integrated for enhanced multi-scale feature extraction.

Main Results:

• The Bi-DCNet model achieved a Dice Similarity Coefficient (DSC) of 0.9228 and an Intersection over Union (IoU) of 0.8576.
• The model demonstrated effectiveness on the EchoNet-Dynamic dataset, a large clinical video dataset.
• This marks the first implementation of a bilateral-structured network for left ventricular segmentation on this dataset.

Conclusions:

• The proposed Bi-DCNet model provides an effective solution for automated left ventricular segmentation in echocardiography.
• The integration of spatial and contextual paths with dilated convolutions improves segmentation accuracy.
• The study highlights the potential of advanced deep learning architectures for cardiac image analysis.