ARCHIVES
VOL. 10, ISSUE 4 (2025)
An optimized convolution neural network technique for stroke detection in magnetic resonance images
Authors
Atinuke Omowumi Aderinto, Justice Ono Emuoyibofarhe, Stephen Olatunde Olabiyisi
Abstract
Artificial Intelligence (AI) is the use of
mathematical algorithms to mimic human cognitive abilities. Convolution Neural
Network (CNN) is an outstanding branch of deep learning applications in
artificial intelligence which could be used in medical imaging specially to
detect neurovascular disease which is stroke in Magnetic Resonance images. It has been difficult to train considerably deep CNNs with the lack of
training data, the poor spatial resolution of MR images and the need for a
short prediction time. However, existing CNN
based systems are limited by inability to perform faster convolutions thereby
rendering them inadequate to efficiently manage patients with neurovascular
diseases like stroke. Hence this work developed an optimized convolution neural
network (CNN-GSA) technique that detect neurovascular disease (stroke) in
magnetic resonance images. 5000 datasets were retrieved from
kaggle which is an online public repository. The retrieved datasets were
pre-processed by employing median filter and image enhancement techniques. The
pre-processed images were segmented by using garbor filter technique.
Gravitational Search Algorithm (GSA) was used to assign optimal weight
parameters for the Convolution Neural Network (CNN) to produce CNN-GSA
technique. The developed method (CNN-GSA) was used for feature extraction and
detection. The implementation of the CNN-GSA was done by using Matlab R2020a.
The performance of the CNN-GSA was evaluated based on Precision, Sensitivity,
Specificity, False Positive Rate, Recognition Accuracy and computational time.
There are eight (8) classifications for the model’s predictions of True
Positive (TP), False Negative (FN), False Positive (FP) and True Negative (TN).
The results showed that the standard CNN based system has a degree of average
of False Positive Rate (FPR) to be 7.83%, recognition accuracy to be 92.57%,
sensitivity which is also recall to be 92.98%, precision to be 92.24%,
specificity to be 92.17% and computational time to be 86.98% across all
classes. The binary threshold across all classes were 0.25, 0.25, 0.35, 0.5,
0.5, 0.75 and 0.75 respectively. The result of the developed CNN-GSA has a
degree of average of False Positive Rate (FPR) to be 3.15%, Recognition
Accuracy to be 97.26%, sensitivity which is also recall to be 97.67%, precision
to be 96.89%, specificity to be 96.85% and average computational time to be
54.59% seconds across all classes. It
is evident from the results that Convolution Neural Network algorithm can be
used to detect neurovascular disease (stroke) in magnetic resonance images but
an optimized Convolution Neural Network (CNN–GSA) technique is needed to assist
radiologists in making the right decisions with minimum errors. Hence
Radiologist can use the developed system as a second opinion, which will
greatly improve diagnosis of neurovascular disease (stroke) in order to reduce
stroke mortality rate.
Download
Pages:33-37
How to cite this article:
Atinuke Omowumi Aderinto, Justice Ono Emuoyibofarhe, Stephen Olatunde Olabiyisi "An optimized convolution neural network technique for stroke detection in magnetic resonance images". International Journal of Advanced Scientific Research, Vol 10, Issue 4, 2025, Pages 33-37
Download Author Certificate
Please enter the email address corresponding to this article submission to download your certificate.