Skip to content

ashwinkumargb/pediatric_gmseg

Repository files navigation

Pediatric_GMSeg

Transfer learning to Improve Pediatric Spinal Cord Gray Matter (GM) segmentation. Final project for CS 5267 (Deep Learning, Spring 2022).

Summary:

Current spinal cord GM segmentation methods are trained using adult spinal cord datasets and sub-optimally translate to the pediatric population. This project aims to improves pediatric GM segmentation by conducting transfer learning on a pretrained deep learning model (sct_deepseg_gm : https://arxiv.org/abs/1710.01269) using a clinical pediatric spinal cord dataset. After model evaluation, it was shown that transfer learning does improve pediatric gray matter segmentation.

The report that details the experiment in more detail can be found here.

model

Training Overview

The following models and data combinations were trained: (1) Pretrained model without data augmentation, (2) Model (no pretraining) without data augmentation, (3) Pretrained model with data augmentation, and (4) Model (no pretraining) with data augmentation. The models were trained on the Vanderbilt ACCRE high-performance cluster and the scripts can be found in the accre_scripts directory. The python scripts were designed to take advantage of GPU computing and parallelization for efficiency. ACCRE requires the following python dependency versions to run: Python (v3.6.3), Tensorflow (v1.8.0), Keras (v2.20), CUDA (v11.10).

The datasets used for the models are publicly available here.

Repository Structure Explained

Base Directory

GM_Models Directory

Accre_Scripts Directory

Model_Eval Directory

  • model_eval_combined.ipynb: Evaluate the different models and plot the testing results from them accordingly. Models were evaluated using various performance metrics.

About

Pediatric Spinal Cord Gray Matter Segmentation

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published