Skip to content

This repository contains the implementation of the methods presented in the paper "Effective semantic segmentation in Cataract Surgery: What matters most?" to be presented at MICCAI 2021. https://arxiv.org/pdf/2108.06119

License

Notifications You must be signed in to change notification settings

RViMLab/MICCAI2021_Cataract_semantic_segmentation

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Effective Semantic Segmentation in Cataract Surgery: What matters most?

Presented at MICCAI 2021

fig

Effective Semantic Segmentation in Cataract Surgery: What matters most?,
Theodoros Pissas*, Claudio S. Ravasio*, Lyndon Da Cruz, Christos Bergeles (* equal contribution)

arXiv technical report (arXiv 2108.06119)

MICCAI 2021 (proceedings)

News

[28/09/2021] Paper to be presented at MICCAI 2021

Introduction

This repository contains the implementation and model weights of the methods presented in the paper "Effective semantic segmentation in Cataract Surgery: What matters most?", presented at MICCAI 2021.

Abstract

Our work proposes neural network design choices that set the state-of-the-art on a challenging public benchmark on cataract surgery, CaDIS. Our methodology achieves strong performance across three semantic segmentation tasks with increasingly granular surgical tool class sets by effectively handling class imbalance, an inherent challenge in any surgical video. We consider and evaluate two conceptually simple data oversampling methods as well as different loss functions. We show significant performance gains across network architectures and tasks especially on the rarest tool classes, thereby presenting an approach for achieving high performance when imbalanced granular datasets are considered.

PWC

PWC

PWC

Data and requirements

  1. CaDIS dataset can be found here here

  2. Create conda environment with pytorch 1.7 and CUDA 10.0

    conda env create -f environment.yml 
    conda activate SemSegCat

Train

To train a model we specify most settings using json configuration files found in configs.

For training with repeat factor sampling and Lovasz on task 2:

  • For OCRNet :
    python main.py -c configs/OCRNet_rf_lvsz.json --task 2 -d 0 --data_path "path_to_data/segmentation"
  • For DeepLabv3Plus training with repeat factor sampling and Lovasz on task 2:
    python main.py -c configs/DeepLabv3Plus_rf_lvsz.json --task 2 -d 0 --data_path "path_to_data/segmentation"

For running on tasks 1 or 3 specify --task and --task 1 or --task 3 respectively in the filenames.

For setting training batch size to N add command line argument --batch_size N

For modifying other settings such as epochs, augmentation, learning rate etc please modify configuration file.

Applying blacklisting and/or using relabelled data

We identified significantly mislabelled images in the dataset which we exclude. We refer to this as blacklisting . For a small part of those images we also provide corrected labels. Part of the results in the paper is reported on this filtered version of the dataset. The ids of the frames that can be blacklisted can be found in data/data.csv.

  • To apply blacklisting:
    python main.py -c path_to_configuration --data_path "path_to_data/segmentation" --blacklisting True
  • To use the 40 relabelled images please move the directory relabelled inside the your CaDIS data directory at data_path/CADIS/segmentation/. Then run
    python main.py -c path_to_configuration --data_path "path_to_data/segmentation" --use_relabeled True
  • To do both of the above:
    python main.py -c path_to_configuration --data_path "path_to_data/segmentation" --blacklisting True --use_relabeled True

These two options can be used in for both train and inference.

Run a pretrained model on the test set

We provide model weights of our top-performing model on all 3 tasks of the CaDIS dataset.

  1. Download checkpoint directories using from here
  2. Move the downloaded directories inside logs. Note: to use the pretrained models the directory names must NOT be modified.
  3. For test set inference on task 1,2,3 run
    python main.py -c configs/OCRNet_pretrained_t1.json -d 0 --data_path "path_to_data/segmentation"
    python main.py -c configs/OCRNet_pretrained_t2.json -d 0 --data_path "path_to_data/segmentation"
    python main.py -c configs/OCRNet_pretrained_t3.json -d 0 --data_path "path_to_data/segmentation"

Pretrained Models

Pretrained models can be found here

Trained on CaDIS Dataset train-val-test split

(no blacklisting nor relabelled data used)

Task Backbone Model Loss Sampling Test set mIoU CKPT args to train args for inference
1 ResNet-50 OCR Lovasz Repeat Factor 86.40 ckpt -c configs/OCRNet_rf_lvsz.json -t 1 -c configs/OCRNet_pretrained_t1.json
2 ResNet-50 OCR Lovasz Repeat Factor 79.40 ckpt -c configs/OCRNet_rf_lvsz.json -t 2 -c configs/OCRNet_pretrained_t2.json
3 ResNet-50 OCR Lovasz Repeat Factor 71.94 ckpt -c configs/OCRNet_rf_lvsz.json -t 3 -c configs/OCRNet_pretrained_t3.json

The provided models achieve SOTA performance as shown in the table below:

fig

Qualitative results

Qualitative results on unseen surgical video can be found here.

Licensing and copyright

Please see the LICENSE file for details.

Acknowledgements

This work was supported through an Invention for Innovation grant [II-LB-0716-20002] by the National Institute for Health Research, through a grant [714562] by the European Research Council and by Sir Michael Uren Foundation.

Citation

If you found this code useful please cite the following work:

@InProceedings{10.1007/978-3-030-87202-1_49,
author="Pissas, Theodoros
and Ravasio, Claudio S.
and Da Cruz, Lyndon
and Bergeles, Christos",
title="Effective Semantic Segmentation in Cataract Surgery: What Matters Most?",
booktitle="Medical Image Computing and Computer Assisted Intervention -- MICCAI 2021",
year="2021",
pages="509--518",
isbn="978-3-030-87202-1"
}

About

This repository contains the implementation of the methods presented in the paper "Effective semantic segmentation in Cataract Surgery: What matters most?" to be presented at MICCAI 2021. https://arxiv.org/pdf/2108.06119

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages