[Paper Review] P-Mamba: Marrying Perona Malik Diffusion with Mamba for Efficient Pediatric Echocardiographic Left Ventricular Segmentation

Zi Ye, Tianxiang Chen+3 more|arXiv (Cornell University)|Feb 13, 2024

Cardiovascular Function and Risk FactorsMedicine3 citations

TL;DR

P-Mamba proposes a novel, efficient deep learning framework for pediatric echocardiographic left ventricular segmentation by fusing vision Mamba for global dependency modeling with DWT-based Perona-Malik Diffusion (PMD) blocks for noise suppression and edge preservation. The model achieves state-of-the-art accuracy (Dice 0.9221 on PSAX, 0.9046 on A4C) and superior inference efficiency, outperforming vision transformers with quadratic and linear complexity.

ABSTRACT

In pediatric cardiology, the accurate and immediate assessment of cardiac function through echocardiography is crucial since it can determine whether urgent intervention is required in many emergencies. However, echocardiography is characterized by ambiguity and heavy background noise interference, causing more difficulty in accurate segmentation. Present methods lack efficiency and are prone to mistakenly segmenting some background noise areas, such as the left ventricular area, due to noise disturbance. To address these issues, we introduce P-Mamba, which integrates the Mixture of Experts (MoE) concept for efficient pediatric echocardiographic left ventricular segmentation. Specifically, we utilize the recently proposed ViM layers from the vision mamba to enhance our model's computational and memory efficiency while modeling global dependencies.In the DWT-based Perona-Malik Diffusion (PMD) Block, we devise a PMD Block for noise suppression while preserving the left ventricle's local shape cues. Consequently, our proposed P-Mamba innovatively combines the PMD's noise suppression and local feature extraction capabilities with Mamba's efficient design for global dependency modeling. We conducted segmentation experiments on two pediatric ultrasound datasets and a general ultrasound dataset, namely Echonet-dynamic, and achieved state-of-the-art (SOTA) results. Leveraging the strengths of the P-Mamba block, our model demonstrates superior accuracy and efficiency compared to established models, including vision transformers with quadratic and linear computational complexity.

Motivation & Objective

• Address the challenge of noise interference and low efficiency in pediatric echocardiographic left ventricular (LV) segmentation.
• Improve segmentation accuracy by suppressing background noise while preserving fine LV boundary details.
• Enhance computational and memory efficiency compared to existing vision transformer-based models.
• Integrate selective state space modeling and diffusion-based denoising into a unified architecture for ultrasound image segmentation.
• Achieve superior performance on pediatric echocardiography datasets with minimal computational cost.

Proposed method

• Employ a dual-encoder architecture: a vision Mamba encoder for global context modeling with linear complexity and hardware efficiency.
• Introduce DWT-based PMD blocks that apply Perona-Malik diffusion to feature maps after discrete wavelet transform (DWT) decomposition to suppress noise and enhance edges.
• Use DWT to decompose input features into subbands, enabling localized noise filtering while preserving structural details in the LV region.
• Combine feature maps from both Mamba and PMD branches at multiple scales before passing them through a segmentation head and FCN decoder for mask prediction.
• Train the model with multi-scale supervision using Dice loss and other standard segmentation losses to improve boundary accuracy.
• Leverage the Mamba block’s selective structured state space mechanism to model long-range dependencies efficiently without self-attention.

Experimental results

Research questions

• RQ1 Can a hybrid architecture combining Mamba and Perona-Malik diffusion improve segmentation accuracy on noisy pediatric echocardiograms?
• RQ2 Does the integration of DWT-based PMD blocks enhance edge preservation and noise suppression compared to standard CNNs or attention mechanisms?
• RQ3 To what extent does the Mamba-based encoder reduce computational cost while maintaining or improving performance over vision transformers?
• RQ4 How does the proposed P-Mamba model compare in efficiency and accuracy to state-of-the-art models on pediatric echocardiography benchmarks?
• RQ5 What is the contribution of each component (Mamba vs. PMD) to the final segmentation performance, as measured by ablation studies?

Key findings

• P-Mamba achieves a Dice score of 0.9221 on the PSAX dataset and 0.9046 on the A4C dataset, outperforming all compared SOTA methods.
• The model reduces inference time to 23.49 ms, with only 183.37M parameters and 12.22 GB GPU memory usage, significantly outperforming models like PVT and MaxViT in efficiency.
• Ablation studies confirm that removing the DWT-based PMD block reduces Dice by 0.0028 on PSAX, proving its critical role in noise suppression and edge fidelity.
• Replacing the Mamba block with ViT-based models (e.g., PVT, Flatten Transformer) results in lower accuracy, demonstrating Mamba’s superior modeling capability.
• The DWT-based PMD block outperforms a simple Sobel operator, which lacks noise suppression, confirming the necessity of the diffusion mechanism.
• The model maintains high precision (0.9316) and recall (0.9128) on PSAX, indicating robustness to false positives and missed detections.