Skip to content

kyuchoi/dce2ktrans

Repository files navigation

dce2ktrans

This is an official code for "Deep Learning Enhances the Reliability of Dynamic Contrast-Enhanced MRI in Diffuse Gliomas: Bypassing Post-processing and Providing Uncertainty Maps", published in European Radiology.

Summary

A spatiotemporal deep learning model can improve the reliability of DCE-MRI by bypassing the estimation of the AIF and providing uncertainty maps, without diagnostic performance in diffuse glioma grading.

Model's Architecture

Model's architecture

  1. Designed a Deterministic Temporal CNN model (1st stage) to extract physical information variations along the temporal axis.

  2. Developed a Probabilistic model (2nd stage) for the spatial axis to generate mean maps and uncertainty maps for image transformation.

  3. Achieved high-accuracy PK parametric map transformation and provided uncertainty maps to assist radiologists in cancer diagnosis.

Visualization of Synthesis map

Training

  1. Training 3 Channel Outputs of TCN
  • python train_tcn_3ch_fusedlabel.py
  1. Training Probabilistic U-Net model
  • python

To be updated

Inference

  1. Synthesize 3 channel outputs of TCN (same code with train. inference right after train.)
  • python train_tcn_3ch_fusedlabel.py
  1. Synthesize Probabilistic Samples from outputs of TCN
  • python inference_tcn_prob_unet_sample.py
  1. Synthesize Mean, Uncertainty map
  • python inference_mean_uncertainty.py

Dataset

The dataset was temporally split so that the test set (102 patients) consisted of scans taken after March 2016. The remaining 219 patients were randomly split into the training set (165 patients) and the validation set (62 patients).

Data Processing

  • Conversion: All DICOM files converted to NIfTI gzipped files.
  • Corrections: Motion and N4 bias field corrected.
  • DCE Image: Time-averaged DCE image skull stripped and resampled to 256x256 resolution in the xy-plane.
  • Tumor Segmentation:
    • T1WI, T2WI, FLAIR images skull stripped using BET package.
    • Images centered, re-oriented (RAS+), and resampled to 1mm isotropic voxels.
    • T2WI and FLAIR images registered to T1WI space using ANTs SyN algorithm.
    • HD-GLIO neural network used for tumor segmentation.
    • Tumor segmentation maps registered to DCE space using rigid affine transform from ANTs package.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages