|
Title:
|
CONSTRUCTING A STROKE DIAGNOSIS AND PROGNOSIS SYSTEM BASED ON THE BPN ALGORITHM USING TC-99M-ECD SPECT IMAGES |
|
Author(s):
|
Jui-Jen Chen,Hung-Nien Chang Chien and Yen-Hsiang Chang |
|
ISBN:
|
978-989-8704-50-4 |
|
Editors:
|
Piet Kommers, Mário Macedo, Guo Chao Peng and Ajith Abraham |
|
Year:
|
2023 |
|
Edition:
|
Single |
|
Keywords:
|
Stroke Diseases, Cerebral Perfusion Image, Transfer Learning, Back-Propagation Neural Network, Nuclear Medical |
|
Type:
|
Poster |
|
First Page:
|
393 |
|
Last Page:
|
395 |
|
Language:
|
English |
|
Cover:
|
|
|
Full Contents:
|
if you are a member please login 
|
|
Paper Abstract:
|
Background: The technology of deep learning in artificial intelligence (AI) is increasingly applied to medical image recognition. These applications are focused mainly on the analysis of CT or MRI images and seldom intended to support research of nuclear medical images. SPECT is one of the few examination tools that can sensitively reflect abnormalities in the brain at an early stage. In addition to diagnosing cerebrovascular diseases, it can also be used as a tool for prognostic evaluation. Therefore, constructing a stroke image recognition system for diagnosis and prognosis is the goal of this study. Method: We collected all the Tc-99m ECD SPECT brain images from Kaohsiung Chang Gung Memorial Hospital over a period of five years from 2017 to 2021. A total of 144 medical records that met the ICD-10-Code I60-I69 cerebrovascular disease extraction rules were obtained. In the preprocessing of data, noise and defective images were removed. Data augmentation technology was exploited to avoid overfitting and underfitting due to the small amount of data and obtain more new images for higher generalization ability. The back-propagation neural network (BPN) algorithm was adopted to train stroke images and extract important features according to the distribution of blood flows in the brain. Result: The proposed model is compared with the VGG16 through transfer training. It delivers an accuracy of 94.4% (1.3% higher) and a sensitivity of 90.3% (5.4% higher). Its recall rate and F-score reach 90.3% and 94.2% respectively. The ROC curve and average AUC (0.89±0.08) indicate that this model has an excellent discrimination capability. Conclusion: Based on the strengths of the BPN algorithm, we construct a stroke image recognition system to support stroke image recognition, assessment of the possibility of a second stroke, and prognostic evaluation. This system can also assist physicians in performing a rapid diagnosis and reduce errors. It is hoped that the developed software can be ported to real-world medical systems for testing, so as to connect the theory with practical situations. |
|
|
|
|
|
|
