yuukias / heart_pipeline

An enhanced version of Wenjia Bai's Heart Pipeline

Overview

This is a modified version of Wenjia Bai's Pipeline to process UKBiobank data.

Preparation

Please make sure data_raw_dir in config.py have the following structure:

├── 20205
├── 20205_retest
├── 20207
├── 20208
├── 20208_retest
├── 20209
├── 20209_retest
├── 20210
├── 20210_retest
├── 20211
├── 20212
├── 20213

where retest should be the same as retest_suffix field in config.py.

heart_pipeline's Issues

Implement features that require both LV and LA

Finish all strain features

Also need to try whether torsion can be implemented.

LA Strain Rate

Need to check whether it can be implemented.

Refer to Quantification of left atrial strain and strain rate using Cardiovascular Magnetic Resonance myocardial feature tracking: a feasibility study

Implement some LA basic features

• Problem1: PER too large if I use correct lowess function.
• Problem2: Pre-contraction volume might be a bit large, resulting in large passive EF

Run LV features in pipeline

Incorporate features in explainable paper

Validate the pipeline: Compare with existing results and get ICC for revisits

Extract all others features in UKBiobank such as blood pressure

Implement LA strain features

Incorporate UUMamba and nnUNet, evaluate performance of each segmentation model

UUMamba has the problem of vanishing Dice and Loss.

I will continue work on nnUNetV2

Incorporate ECG features to Pipeline

Improve the registration when calculating strain

Currently mirtk compose-dofs -approximate will cause the registration to fail. However, if we do not use approximation, then average_3d_ffd will fail.

In addition, the dice score for Myo is not high for both long axis and short axis.