ZMIC Journal Club

PolarNet for VT Screening


Automated Characterization of Myocardial Scar Topological Patterns for Ventricular Tachycardia Screening


张杨
SDS, Fudan University
2025-07-04

VT Screening
ZMIC Journal Club

Paper Info

  • Accepted by MICCAI 2025 (ratings from reviewers: 5,5,3)
Intro
ZMIC Journal Club

Table Of Content

  • Introduction
    • VT Screening
    • Scar Projection
  • Related Work
    • Scar Segmentation
    • Boundary Aware Segmentation
  • Methods
  • Experiments
  • Extensions
Intro
ZMIC Journal Club

VT Screening

  • Ventricular Tachycardia: your heart beats too fast

  • Causes: Cardiomyopathy(心肌病), Structural heart disease(结构性), Ischemic heart disease(缺血性), Heart failure(心衰), Myocarditis(心肌炎) and so on

  • Treatment: Implantable cardioverter defibrillator(ICD)

Intro
ZMIC Journal Club

VT Screening with CMRI

Our collaborator found that scar projection and classification from LGE CMRI is valuable for VT Screening.

Intro
ZMIC Journal Club

Scar Projection & Classification

In polar-coordinate system, the projection can be easier to understand:

Scar segmentation is necessary in this process and its boundary quality is very important.

Intro
ZMIC Journal Club

Scar Segmentation

Numerous methodologies have been adapted for this task:

  • Traditional: thresholding, region growing
  • Machine learning:SVMs, CRFs
  • Deep learning: UNet-based, Transformer-based

Among them, nn-UNet is a popular SOTA framework:

Related Work
ZMIC Journal Club

Boundary Aware Segmentation?

Boundary quality in scar segmentation has received limited attention. Existing methods still struggle to localize scars within the myocardium, largely due to the irregular morphology and blurred boundaries of pathological scars.

Related Work
ZMIC Journal Club

Retina OCT Boundary Regression

  • He, Yufan, et al. "Structured layer surface segmentation for retina OCT using fully convolutional regression networks." Medical image analysis 68 (2021): 101856.

Related Work
ZMIC Journal Club

Methods overview

Methods
ZMIC Journal Club

Methods details: polar transform

  • Polar transformation and its inverse transformation: implement with torch.nn.functional.grid_sample
    • polar mesh grid:

  • All results are evaluated in Cartesian coordinate system.
Methods
ZMIC Journal Club

Methods details: dual branch

  • Dual branch: boundary & region
    • channel-wise softmax --> -->
    • column-wise softmax --> -->
Methods
ZMIC Journal Club

Methods details: loss function

  1. Loss for :

    Where , is the ground-truth radial position of boundary k at angle θ.

  2. Loss for
    Where is the final boundary position for boundary, .

Methods
ZMIC Journal Club

Experiments: Scar Classfication

Experiments
ZMIC Journal Club

Experiments: Scar Segmentation

Experiments
ZMIC Journal Club

Experiments: Case study

Experiments
ZMIC Journal Club

Extensions

  • Semi-supervised learning(polar + Cartesian)

Extensions
ZMIC Journal Club

THANKS

THANKS