Semantic Memory Guided Image Representation for Polyp Segmentation (ICASSP-2023)

Introduction

The repository contains the PyTorch implementation of "Semantic Memory Guided Image Representation for Polyp Segmentation" ,ICASSP-2023.

Overview

1. Framework

Figure 1: Overview of our proposed DCRNet

2. Quantitative Results

Quantitative results on three benchmarks

Method	Year	EndoScene			Kvasir-SEG			PICCOLO
		MAE	Dice	IoU	MAE	Dice	IoU	MAE	Dice	IoU
U-Net	2015	4.4	73.78	66.54	4.2	85.97	78.70	5.0	66.81	60.59
PraNet	2020	3.5	81.73	74.38	3.1	89.20	83.61	3.0	75.34	69.77
ACSNet	2020	3.0	85.15	78.67	3.2	89.32	83.83	2.2	79.08	74.82
SCRNet	2021	2.9	85.29	78.81	3.6	88.62	82.52	3.0	78.51	72.74
SANet	2021	3.1	84.15	77.21	3.1	89.86	79.04	3.6	79.08	73.04
CCBANet	2021	3.4	83.85	76.52	3.1	89.43	83.41	2.8	76.22	71.64
MSNet	2021	3.6	80.82	74.49	3.4	89.03	83.08	3.2	81.37	75.58
Ours	2022	2.9	86.03	79.34	3.2	89.92	83.93	2.8	83.04	76.78

3. Qualitative Results

Figure 2: Qualitative results of different methods on PICCOLO.

Usage

1. Prerequisite environment

1. torch>=1.5.0

2. torchvision>=0.6.0

3. tqdm

4. scipy

5. scikit-image

6. PIL

7. numpy

8. CUDA

2. Dataset downloading

• Downloading the CVC-EndoSceneStill dataset, which can be found in this Google Drive link
• Downloading the Kvasir-SEG dataset, which can be found in this Google Drive link
• To access the PICCOLO dataset, please visit here

3. Train

• Assign your customized path of --train_path ,--save_root and --gpu in Train.py.
• Run python Train.py

4. Test

• Assign the --pth_path , --data_root , --save_root and --gpu in Test.py.
• Run python Test.py
• The quantitative results will be displayed in your screen, and the qualitative results will be saved in your customized path.

5. Evaluate

• The evaluation code is stored in ./utils/eval.py
• You can replace it with your customized evaluation metrics.