Abstract Background/Introduction The left ventricle is vital for assessing heart function, and its analysis in echocardiographic images supports the diagnosis of many cardiac conditions. However, accurate segmentation is difficult due to image noise, low contrast, and anatomical variability. Using anatomical information and prior shape knowledge can improve the robustness and accuracy of segmentation, especially in challenging cases. Purpose Deep learning techniques are widely used for cardiac segmentation due to their strong performance in image analysis. This work aims to enhance segmentation robustness by integrating prior shape information directly into the network, allowing it to better handle challenging image conditions. Method This shape information can be obtained from deep statistical shape models (SSM) that capture the variability of shapes within a population as a statistical representation. We present a fusion-based segmentation approach that combines a state-of-the-art segmentation model with shape information from deep SMM. This shape information is incorporated into the training process through an additional loss term. In our method, we use U-Net for segmentation and employ Base-DeepSSM 2 (Fig. 1b) and TL-DeepSSM 3 (Fig. 1c) frameworks to extract shape information in the form of latent codes. These codes are generated using frozen encoders from pre-trained Base/TL-DeepSSM models. The added loss function (Lₗatent) measures the difference between the latent codes of the predicted segmentations and the original input images (Fig. 1ₐ). The proposed loss function helps maintain the anatomical shape in the predicted segmentation map. The final loss is the weighted sum of latent loss and dice loss. Results We applied our proposed method to the Echocardiography Analysis (MITEA) dataset 1. In each image, the endocardium and epicardium of the left ventricle are manually labeled, providing precise anatomical segmentation for training and evaluation. A selected subset of the data was divided into 54 image-label pairs for training (75%), 12 for validation (17%), and 14 for testing (8%). The training runs for a total of 60 epochs: the first 20 epochs train only the U-Net using Dice Loss, and the following 40 epochs include both Dice Loss and the latent loss. The proposed model achieved a highest validation Dice score of 0. 84 and reached a Dice of 0. 83 on the test dataset for epicardium, with a standard deviation of 0. 07. In comparison, the baseline U-Net model obtained a validation Dice score of 0. 81 and a test Dice of 0. 80, with a higher standard deviation of 0. 10. These results indicate that the fusion model offers improved segmentation accuracy and greater consistency over the standard U-Net. A promising direction for future work is to embed shape information directly into the network architecture, such as the U-Net’s encoder, instead of using it solely as an additional loss function. Block diagram of the proposed methodQualitative Results
Yildirim et al. (Thu,) studied this question.