syanng/blazeneo

BlazeNeo: Blazing fast polyp segmentation and neoplasm detection

This repository provides code for paper "BlazeNeo: Blazing fast polyp segmentation and neoplasm detection" in IEEE Access, vol. 10, pp. 43669-43684, 2022.

Features

• Real-time Speed: Our method, name BlazeNeo, can efficiently perform polyp segmentation and neoplasm detection from polyp videos with real-time speed (~81.5 FPS) on a single NVIDIA Tesla V100 GPU.
• Cutting-edge Performance: Extensive experiments on the newly collected NeoPolyp dataset and comparisons to existing models. Moreover, we measure model latency and throughput on dedicated hardware in a setting similar to a real-life deployment.

Framework overview

Tutorials

Installation

Download our public dataset BKAI-IGH NeoPolyp-Small for your own experiments.

pip install -r requirements.txt

Model training

python train.py

Model evaluation

python eval.py

Prediction

python infer.py

Quantitative comparison

Performance metrics on the NeoPolyp-Clean test set for ColonSegNet, U-Net, DDANet, DoubleU-Net, PraNet, HardNet-MSEG, NeoUNet and BlazeNeo.

Method	Diceseg	Dice non	Dice neo	FPS	Parameters	GFLOPs
ColonSegNet	0.738	0.505	0.732	44.9	5,010,000	64,84
U-Net	0.785	0.525	0.773	69.6	31,043,651	103,59
DDANet	0.813	0.578	0.802	46.2	6,840,000	31,45
DoubleU-Net	0.837	0.621	0.832	43.2	18,836,804	83,62
HarDNet-MSEG	0.883	0.659	0.869	77.1	17,424,031	11,38
PraNet	0.895	0.705	0.873	55.6	30,501,341	13,11
NeoUNet	0.911	0.720	0.889	68.3	38,288,397	39,88
BlazeNeo (Ours)	0.904	0.717	0.885	81.5	17,143,324	11.06

Qualitative comparison

Qualitative comparison of the proposed method with other baseline methods: (a) image, (b) ground truth, (c) BlazeNeo (Ours), (d) NeoUNet, (e) PraNet, (f) HarDNet-MSEG, (g) UNet, (h) Double U-Net, (i) DDANet, and (j) ColonSegNet.

Ackowledgement

This work was funded by Vingroup Innovation Foundation (VINIF) under project code VINIF.2020.DA17. We thank IGH for collecting and annotating the data.

Cite

Please cite our work if you find it useful:

@ARTICLE{blazeneo2022,
  author={An, Nguyen S. and Lan, Phan N. and Hang, Dao V. and Long, Dao V. and Trung, Tran Q. and Thuy, Nguyen T. and Sang, Dinh V.},
  journal={IEEE Access}, 
  title={BlazeNeo: Blazing Fast Polyp Segmentation and Neoplasm Detection}, 
  year={2022},
  volume={10},
  pages={43669-43684}
}

FAQ

If you want to improve the usability or any piece of advice, please feel free to contact me directly ([email protected])

blazeneo's Issues

Issues related to training code

def main(argv):
YAML_FP = "/home/s/polyp/neoplasm-segmentation/config/exp190122.yml"
SAVED_FN = "DUNet_190122.pth"

with open(YAML_FP, "r") as fp:
    config = yaml.load(fp)

This section of the training code puzzles me. What is "/home/s/polyp/neoplasm-segmentation/config/exp190122.yml" and what is the "DUNet" that appears?
Thank you for your attention and answers.