amanbasu / 3d-prostate-segmentation

Segmentation of prostate from MRI scans
GNU General Public License v3.0
44 stars 13 forks source link
deep-learning fully-convolutional-networks medical-imaging mri-images python tensorflow

3D Prostate Segmentation from MR Images using FCNN

Collaborator: Aman Agarwal, Aditya Mishra

This repository contains files related to Volumetric Segmentation of Prostate from MR Images Using FCNN with Increased Receptive Field, presented at Nvidia GTC 2019 (link). The dataset is provided by PROMISE12 challenge.

DOI 10.1134/S1054661821020024

About the files

Introduction

Data

Implementation

We used a modified V-net architecture for segmentation shown in the figure below.

Training

We trained our model on different GPUs and got the following speedups.

GPU configuration Batch Size Average Time per Epoch (s)
Single K80 2 147
Dual k80 2 (1 per GPU) 102
Single P100 2 48
Single P100 5 27

Evaluation Metrics

The metrics used in this study are widely used for the evaluation of segmentation algorithms:

  1. Dice coefficient: To measure the similarity between output volumes.
  2. Absolute relative volume difference: the percentage of the absolute difference between the volumes.
  3. Average boundary distance: the average over the shortest distances between the boundary points of the volumes.
  4. 95% Hausdorff distance: the maximum of the shortest distances between the boundary points of the volumes. 95% percentile makes the metric less sensitive to outliers.

Results

After training for 5700 epochs we got a dice loss of 0.94 and 0.87 on training and validation set. The results were then submitted to the MICCAI PROMISE12 challenge, and we received a score of 84.61 on the test set.