[Paper Review] Automated Segmentation of Left Ventricle in 2D echocardiography using deep learning

Neda Azarmehr, Xujiong Ye+4 more|arXiv (Cornell University)|Mar 17, 2020

Advanced Neural Network Applications4 references4 citations

TL;DR

This study evaluates the performance of the original U-Net and two advanced U-Net variants (U-Net 1 and U-Net 2) for automated left ventricle segmentation in 2D echocardiography. Despite claims of improved accuracy, the original U-Net outperformed both variants, achieving a mean Dice coefficient of 0.92 ± 0.05 and Hausdorff distance of 3.97 ± 0.82 on an independent test set of 199 frames, demonstrating robustness comparable to expert readers.

ABSTRACT

Following the successful application of the U-Net to medical images, there have been different encoder-decoder models proposed as an improvement to the original U-Net for segmenting echocardiographic images. This study aims to examine the performance of the state-of-the-art proposed models as well as the original U-Net model by applying them to segment the endocardium of the Left Ventricle in 2D automatically. The prediction outputs of the models are used to evaluate the performance of the models by comparing the automated results against the expert annotations (gold standard). Our results reveal that the U-Net model outperforms other models by achieving an average Dice coefficient of 0.92$ \\pm 0.05$, and Hausdorff distance of 3.97$ \\pm 0.82$.

Motivation & Objective

• To evaluate the performance of state-of-the-art U-Net-based models for automated left ventricle segmentation in 2D echocardiography.
• To compare the original U-Net with U-Net 1 and U-Net 2 on an independent dataset to assess generalization and robustness.
• To determine whether newer U-Net variants improve segmentation accuracy over the original U-Net in clinical echocardiographic imaging.
• To investigate the impact of image quality on model performance and compare automated results with expert annotations.

Proposed method

• The study used a dataset of 61 patients with apical 4-chamber 2D echocardiographic views, acquiring 992 annotated frames from end-diastolic and end-systolic phases.
• Ground-truth segmentation was established by an experienced cardiologist, with a second expert providing blinded annotations to assess inter-observer variability.
• The dataset was split into 60% training, 20% validation, and 20% testing, with no patient overlap across sets to ensure generalization.
• All models were trained using the Adam optimizer with a learning rate of 0.00001 and 250 training epochs, using negative log-likelihood loss.
• Input images were resized to 256×256 pixels to ensure fair comparison across models, and all models produced output of the same spatial resolution.
• Model evaluation was performed using the Dice coefficient and Hausdorff distance, comparing automated predictions against expert annotations.

Experimental results

Research questions

• RQ1 Does the original U-Net outperform U-Net 1 and U-Net 2 in segmenting the left ventricle in 2D echocardiography?
• RQ2 How do the performance metrics of automated segmentation models compare to inter-observer variability among expert cardiologists?
• RQ3 To what extent does image quality affect the accuracy of deep learning-based segmentation models in echocardiography?
• RQ4 Can automated models achieve performance comparable to human experts in left ventricle segmentation?

Key findings

• The original U-Net achieved the highest mean Dice coefficient of 0.92 ± 0.05 across the test set, outperforming both U-Net 1 and U-Net 2.
• The original U-Net recorded the lowest Hausdorff distance of 3.97 ± 0.82, indicating superior boundary localization accuracy.
• Visual inspection confirmed that the original U-Net produced smoother and more accurate LV endocardial contours compared to the other models.
• All models showed reduced performance on low-quality images, particularly those with speckle noise, poor contrast, or missing borders.
• The automated model's disagreement with expert annotations was comparable to the disagreement between the two expert readers, indicating robustness.
• The study found no significant improvement in segmentation accuracy from newer U-Net variants over the original U-Net on this dataset.