[Paper Review] Multi-Scale Fully Convolutional Network for Cardiac Left Ventricle Segmentation

Han Sung Kang, Defeng Chen|arXiv (Cornell University)|Sep 19, 2018

Radiomics and Machine Learning in Medical ImagingMedicine11 references3 citations

TL;DR

This paper proposes MS-FCN, a multi-scale fully convolutional network for cardiac left ventricle segmentation that integrates a multi-scale pooling module for enhanced contextual feature extraction and a dense connectivity decoder to refine boundary details. On the Sunnybrook cine-MR dataset, it achieves state-of-the-art performance with Dice scores of 0.93 for the endocardium and 0.96 for the epicardium.

ABSTRACT

The morphological structure of left ventricle segmented from cardiac magnetic resonance images can be used to calculate key clinical parameters, and it is of great significance to the accurate and efficient diagnosis of cardiovascular diseases. Compared with traditional methods, the segmentation algorithms based on fully convolutional neural network greatly improve the accuracy of semantic segmentation. For the problem of left ventricular segmentation, a new fully convolutional neural network structure named MS-FCN is proposed in this paper. The MS-FCN network employs a multi-scale pooling module to ensure that the network maximises the feature extraction ability and uses a dense connectivity decoder to refine the boundaries of the object. Based on the Sunnybrook cine-MR dataset provided by the MICCAI 2009 challenge, numerical experiments demonstrate that our proposed model has obtained state-of-the-art segmentation results: the Dice score of our method reaches 0.93 on the endocardium, and 0.96 on the epicardium.

Motivation & Objective

• To address the challenge of accurate and efficient left ventricle segmentation in cardiac MRI, which is critical for diagnosing cardiovascular diseases.
• To overcome limitations of traditional manual segmentation, which is time-consuming and prone to inter-observer variability.
• To improve semantic segmentation performance by enhancing feature extraction and boundary refinement in deep learning-based methods.
• To achieve state-of-the-art results on the benchmark Sunnybrook cine-MR dataset for left ventricle segmentation.
• To validate the effectiveness of multi-scale pooling and dense connectivity in a fully convolutional encoder-decoder architecture.

Proposed method

• The network employs an encoder-decoder architecture with 15 consecutive 3×3 convolutional layers and three downsampling operations in the encoder stage.
• A multi-scale pooling module replaces standard pooling layers, using deconvolution-based upsampling with learnable parameters to preserve spatial details and enhance contextual representation.
• The decoder uses a dense connection structure to concatenate feature maps from multiple layers, improving gradient flow and feature reuse for boundary refinement.
• Skip connections are incorporated between corresponding encoder and decoder layers to preserve spatial information and support accurate localization.
• The network uses batch normalization and ReLU activation after each convolutional layer to stabilize and accelerate training.
• Group deconvolution with 32 groups is used in the multi-scale pooling module to reduce parameters while maintaining performance.

Experimental results

Research questions

• RQ1 Can a multi-scale pooling module with learnable deconvolutional upsampling improve feature representation in left ventricle segmentation compared to fixed interpolation?
• RQ2 Does a dense connection structure in the decoder significantly enhance boundary precision compared to standard skip connections?
• RQ3 How does the proposed MS-FCN architecture compare to existing state-of-the-art models like FCN and U-Net on the Sunnybrook dataset?
• RQ4 What is the impact of different upsampling methods (bilinear, deconvolution, group deconvolution) on segmentation accuracy and parameter efficiency?
• RQ5 In which anatomical regions (e.g., apex, basal, or ambiguous boundaries) does the model show limitations in segmentation performance?

Key findings

• The proposed MS-FCN achieved a Dice score of 0.93 ± 0.02 for the endocardium and 0.96 ± 0.01 for the epicardium on the Sunnybrook dataset, outperforming previous fully automated methods.
• The use of deconvolution in the multi-scale pooling module improved segmentation performance over bilinear interpolation, with group deconvolution reducing parameters while maintaining accuracy.
• The dense connection decoder reduced average precision distance (APD) to 1.61 mm for endocardium and 1.61 mm for epicardium, outperforming FCN and U-Net in boundary accuracy.
• The percentage of good contours reached 98.35% for endocardium and 98.51% for epicardium, indicating high consistency and reliability in segmentation results.
• Ablation studies confirmed that both the multi-scale pooling module and dense connection structure significantly contributed to performance gains, especially in boundary refinement.
• The model struggled most with ambiguous boundaries and in the apex and basal regions, where image contrast and structure are less distinct.