Abstract
This paper introduces the Efficient Decoupled Masked Autoencoder (EDMAE), a novel self-supervised method for recognizing standard views in pediatric echocardiography. EDMAE introduces a new proxy task based on an encoder-decoder structure where the encoder is decoupled into a teacher and a student component. The teacher encoder extracts representations from visible image blocks, while the student encoder processes masked blocks. A consistency loss between their feature maps ensures robust latent representation learning. By utilizing pure convolution operations instead of Vision Transformers (ViT), EDMAE improves training efficiency and convergence speed. Pre-trained on a large-scale private pediatric dataset and fine-tuned for 27 standard views, EDMAE achieves high classification accuracy and demonstrates superior performance in downstream tasks like cardiac segmentation on the public CAMUS dataset.
Introduction
Identifying standard views in pediatric echocardiography is a critical clinical task, complicated by the anatomical variability inherent in congenital heart diseases (CHDs). While deep learning has shown promise for automated view recognition, the scarcity of expert-annotated medical data remains a significant bottleneck. Self-supervised learning (SSL), particularly Masked Autoencoders (MAE), offers a solution by learning from unlabeled data. However, standard MAEs often rely on computationally intensive Vision Transformers (ViTs) and can struggle with semantic consistency when masking is aggressive.
Methodology
The proposed EDMAE framework addresses these challenges through two primary innovations: architectural decoupling and the use of efficient convolutional backbones.
Teacher-Student Decoupling
The core of EDMAE is its "decoupled" encoder structure. Unlike traditional MAEs that use a single encoder, EDMAE employs a dual-encoder setup:
- Teacher Encoder: Processes the original, visible image blocks to extract a high-fidelity potential representation.
- Student Encoder: Processes the masked version of the same image.
A feature alignment loss is calculated between the feature maps of both encoders. This forces the student encoder to predict the latent features of missing patches, ensuring that the learned representations are semantically consistent regardless of whether specific image regions are visible or masked.
Pure Convolutional Backbone
Moving away from the standard ViT-based MAE, EDMAE utilizes a pure convolutional architecture, specifically DenseNet for the encoder and a lightweight CNN for the decoder. Convolutions are inherently suited for the local spatial patterns in ultrasound images and offer significantly faster convergence and lower computational overhead compared to Transformers on specialized medical datasets.
Proxy Task and Reconstruction
The model is pre-trained by simultaneously optimizing the feature alignment loss and a reconstruction loss. The decoder's task is to reconstruct the original image from the combined feature maps, compelling the encoder to capture essential anatomical structures.
Experimental Results
The effectiveness of EDMAE was evaluated on a private pediatric dataset and the public CAMUS dataset.
Standard View Identification
On a private dataset from the Shanghai Children’s Medical Center containing 27 standard views, EDMAE achieved an overall accuracy of 98.48% and a mean F1-score of 93.63%. These results significantly outperform standard MAE and other SSL methods like MoCoV2 and ConvMAE.
Downstream Segmentation
To verify generalizability, EDMAE was applied to cardiac ultrasound segmentation on the CAMUS dataset. It achieved a 93.09% Dice Coefficient, surpassing specialized supervised models like TransUNet, demonstrating the robustness of its learned features.
Training Efficiency
A key advantage of the convolutional approach is efficiency. EDMAE reached convergence in approximately 24 hours, a 3x speedup compared to the 72 hours required by standard ViT-based MAE architectures.
Conclusion
EDMAE presents an efficient and effective framework for self-supervised learning in pediatric cardiology. By decoupling the encoder into a teacher-student pair and leveraging convolutional architectures, the model achieves state-of-the-art performance in view identification and segmentation while drastically reducing training time. Future work will explore the application of this decoupled approach to other medical imaging modalities and multi-modal clinical data.
Summary updated