Table 3 Comparison of different deep learning methods for wound segmentation in different studies.

From: Automatic segmentation and measurement of pressure injuries using deep learning models and a LiDAR camera

Authors (year)

The best model

(Total number of trained models)

Type of ulcer

Database (number)

Internal validation

(The best model)

External validation

(The best model)

Comments

DC

IoU

DC

IoU

Our study, 2022

U-Net (2)

PU

Internal:327

External:201

0.944

0.898

0.849

0.777

1. Combined with two state-of-art models

2. Combined with automatic wound area measurement

Wang et al.10

ConvNet (2)

Chronic ulcers

Internal:500

External:150

N/A

N/A

N/A

0.473

Compared with machine learning (SVM)

Goyal et al.11

FCN-16 s (4)

DFU

Internal:480

External:120

N/A

N/A

0.794

N/A

Compared with FCN with different layers

Liu et al.12

MobileNet-FCN16

(3)

Ulcers

Internal: 900

0.917

0.846

N/A

N/A

1. Lacks external validation

2. Using watershed algorithm

Wang et al.13

MobileNetV2 with CCL (6)

Foot ulcers

Internal: 1109

External: Medetec dataset

0.905

N/A

0.945

N/A

Requiring post-segmentation process

Chang et al.14

DeepLabV3

(5)

PU

Internal: 2893

0.989

0.978

N/A

N/A

1. Combined with tissue segmentation

2. Lacks external validation

García-Zapirain et al.30

3D CNN

PU

Internal: 193

0.92

N/A

N/A

N/A

Lacks external validation

Ohura et al.31

U-Net

(4)

PU → DFU + VLU

Internal: 400 (PU)

External:20 (DFU); 20 (VLU)

0.936

N/A

0.850

N/A

Using PU-trained model to do segmentation of DFU and VLU

Zahia et al.32

CNN

PU

22

N/A

N/A

N/A

N/A

Only for tissue classification

  1. DC Dice coefficient, IoU Intersection over union, PU Pressure ulcer, DFU Diabetic foot ulcers, VLU Venous leg ulcer, CNN Convolutional neural network, U-Net Mask R-CNN; SVM; ConvNet; FCN Fully convolutional networks, SVM Support vector machine, CCL Connected component labelling.
Back to article page