Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Memory-based Ensemble Learning in CMR Semantic Segmentation

Yiwei Liu, Ziyi Wu, Liang Zhong, Lingyi Wen, Yuankai Wu

TL;DR

The study tackles ventricular segmentation in cardiac cine MRI, where end-slice accuracy is clinically critical and traditional methods struggle. It introduces Streaming, a memory-based ensemble that leverages global uncertainty extracted from 3D frame variance to dynamically weight base classifiers at a pixel level, balanced by a joint Dice–Focal loss. A key contribution is the End Coefficient (EC), a metric quantifying end-slice performance, and comprehensive evaluation on ACDC and M&Ms showing near-state-of-the-art overall Dice and clear gains on end slices, with efficient end-to-end training. The approach offers clinically meaningful improvements in patient-specific segmentation by stabilizing end-slice predictions without sacrificing global accuracy, making it suitable for routine CMR analysis.

Abstract

Existing models typically segment either the entire 3D frame or 2D slices independently to derive clinical functional metrics from ventricular segmentation in cardiac cine sequences. While performing well overall, they struggle at the end slices. To address this, we leverage spatial continuity to extract global uncertainty from segmentation variance and use it as memory in our ensemble learning method, Streaming, for classifier weighting, balancing overall and end-slice performance. Additionally, we introduce the End Coefficient (EC) to quantify end-slice accuracy. Experiments on ACDC and M&Ms datasets show that our framework achieves near-state-of-the-art Dice Similarity Coefficient (DSC) and outperforms all models on end-slice performance, improving patient-specific segmentation accuracy.

Memory-based Ensemble Learning in CMR Semantic Segmentation

TL;DR

The study tackles ventricular segmentation in cardiac cine MRI, where end-slice accuracy is clinically critical and traditional methods struggle. It introduces Streaming, a memory-based ensemble that leverages global uncertainty extracted from 3D frame variance to dynamically weight base classifiers at a pixel level, balanced by a joint Dice–Focal loss. A key contribution is the End Coefficient (EC), a metric quantifying end-slice performance, and comprehensive evaluation on ACDC and M&Ms showing near-state-of-the-art overall Dice and clear gains on end slices, with efficient end-to-end training. The approach offers clinically meaningful improvements in patient-specific segmentation by stabilizing end-slice predictions without sacrificing global accuracy, making it suitable for routine CMR analysis.

Abstract

Existing models typically segment either the entire 3D frame or 2D slices independently to derive clinical functional metrics from ventricular segmentation in cardiac cine sequences. While performing well overall, they struggle at the end slices. To address this, we leverage spatial continuity to extract global uncertainty from segmentation variance and use it as memory in our ensemble learning method, Streaming, for classifier weighting, balancing overall and end-slice performance. Additionally, we introduce the End Coefficient (EC) to quantify end-slice accuracy. Experiments on ACDC and M&Ms datasets show that our framework achieves near-state-of-the-art Dice Similarity Coefficient (DSC) and outperforms all models on end-slice performance, improving patient-specific segmentation accuracy.

Paper Structure

This paper contains 18 sections, 6 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Ensemble Learning
  4. Uncertainty for Segmentation
  5. Methodology
  6. Ensemble Strategy
  7. Loss Function
  8. End-to-end Training
  9. Experiments
  10. Setting
  11. Datasets
  12. Pre-processing and Hyperparameters
  13. Metrics
  14. Results
  15. Comparison on ACDC and M&Ms
  16. ...and 3 more sections

Figures (4)

  • Figure 1: Visual differences between traditional ensemble learning and ours.
  • Figure 2: a) shows results from 1UNet + 1D3P, with the x-axis varying UNet's weight. b) shows results from 2UNet, with the x-axis varying UNet 1's weight.
  • Figure 3: Test FLOPs and parameters on a 3D frame, marker size shows parameter count.
  • Figure 4: UNet Trio is 3UNet (Uncertainty), and UNet i is one of its components. Solo means working individually. $-1$ and $-2$ are the last two slices, $0$ and $1$ the first two.