perceivelab / panknet

The official PyTorch implementation for paper "Hierarchical 3D Feature Learning for Pancreas Segmentation"

Hierarchical 3D Feature Learning for Pancreas Segmentation

Federica Proietto Salanitri, Giovanni Bellitto, Ismail Irmakci, Simone Palazzo, Ulas Bagci, Concetto Spampinato

Overview

Novel 3D fully convolutional deep network for automated pancreas segmentation from both MRI and CT scans. The proposed model consists of a 3D encoder that learns to extract volume features at different scales; features taken at different points of the encoder hierarchy are then sent to multiple 3D decoders that individually predict intermediate segmentation maps. Finally, all segmentation maps are combined to obtain a unique detailed segmentation mask. The model outperforms existing methods on CT pancreas segmentation on publicly available NIH Pancreas-CT dataset (consisting of 82 contrast-enhanced CTs), obtaining an average Dice score of about 88%. Furthermore, yields promising segmentation performance on a very challenging private MRI dataset, consisting of 40 MRI scans (average Dice score is about 77%).

Method

Examples

Notes

• As Feature Extractor, PankNet employs S3D pretained on Kinetics-400 dataset. The S3D weights can be downloaded from here.

Citation

@InProceedings{10.1007/978-3-030-87589-3_25,
author="Proietto Salanitri, Federica
and Bellitto, Giovanni
and Irmakci, Ismail
and Palazzo, Simone
and Bagci, Ulas
and Spampinato, Concetto",
editor="Lian, Chunfeng
and Cao, Xiaohuan
and Rekik, Islem
and Xu, Xuanang
and Yan, Pingkun",
title="Hierarchical 3D Feature Learning forPancreas Segmentation",
booktitle="Machine Learning in Medical Imaging",
year="2021",
publisher="Springer International Publishing",
address="Cham",
pages="238--247",
isbn="978-3-030-87589-3"
}

panknet's Issues

help

Hello, what are the environmental requirements for your configuration？

request for a tutorial on using pancreas segmentation

Dear authors,
Thank you for this great work for pancreas segmentation. However, is that possible to add a quick start tutorial on how to use this algorithm for pancreas segmentation? We are indeed need this functionality for our project as we have > 1500 abdo. CT scan for pancreatitis, but we need automatic segmentation of the pancreas.