Table 1 Presents the methods and limitations related to existing studies.

From: Integrating convolutional and transformer networks for precise diagnosis of watershed and hemorrhagic stroke

Article

Dataset size

Class of Ischemic stroke

Segmentation/classification method

Performance

limitation

Advantages

Soltan et al23.

• ISLES 2018

• Training :94 scans,

• testing:62 scan

• acute stroke

• Parallel DNN + pixel level classifier providing advantage of better learning with limited data

• Voting

• simple weighted averaging, logistic regression.

DSC = 71.3%%

Recall = 73.6%%

Volumetric Similarity = 82.1%,

• Requires large number of parameters

• Limited data

• Computation cost missing

• Control overfitting

Luu-Ngoc Do. et al27.

• 71 patients

• acute anterior circulation stroke

• RRCNN

• Accuracy = 87.3%

• Accuracy Method 2= 0.941

• F1score = 0.888

• Low test Accuracy

• Region based approach Missing

• Limited to one class of Ischemic

• Less training time 4 to 5 h as compared to 3DCNN

kumar et al.25

• ISLES 2017

• 28 training and 36 testing

• ISLES 2015

• acute stroke

• CS-Net(U-Net variance)

• DC = 0.84 ± 0.11

• Enough training time 18, 9.5, and

8 h for SISS 2015 SPES 2015

• & ISLES 2017.

• Limited to one class of Ischemic

 

[2]Yi-Chia Wei et al27.

• 216 patients with 4606 slices

• lacune and non-lacune

SGD-net•

Accuracy = 0.806–0.828

• Limitation on small lesion detection

• Scattered lesion detection invisibility

 

[3]De Vries L et al29.

• ISLES 2018

• acute Ischemic Stroke

• PerfU-Net

 

• Require larger learning rate for training

• lack of ground truth penumbra segmentation

• posterior infarcts Missing

 

[4]

Hulin Kuang et al40.

• AISD

dataset

• Acute Ischemic Stroke < 24

• Hybrid convolutional neural

network (CNN) & Transformer network

DC = 61.39%

• Small dataset

 

[5]Xiyue Wang et al.43

• 2836 ICH

• 5 Subtypes of ICH

• CNN-RNN

• Accuracy > 0.8

• Limited Number of subject belong to each class

 
Back to article page