An improved diagnostic algorithm based on deep learning for ischemic stroke detection in posterior fossa
Ischemic stroke is triggered by an obstruction in the blood vessel of the brain, preventing the blood to flow to the brain tissues region. Solving this is extremely beneficial as Non-enhanced Computed Tomography (NECT) has significant shortcomings for posterior fossa (PF): (i) deficient sensit...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English English |
| Published: |
2020
|
| Subjects: | |
| Online Access: | http://eprints.uthm.edu.my/4121/1/24p%20ANIS%20AZWANI%20MUHD%20SUBERI.pdf http://eprints.uthm.edu.my/4121/2/ANIS%20AZWANI%20MUHD%20SUBERI%20COPYRIGHT%20DECLARATION.pdf http://eprints.uthm.edu.my/4121/3/ANIS%20AZWANI%20MUHD%20SUBERI%20WATERMARK.pdf |
| Summary: | Ischemic stroke is triggered by an obstruction in the blood vessel of the brain,
preventing the blood to flow to the brain tissues region. Solving this is extremely
beneficial as Non-enhanced Computed Tomography (NECT) has significant
shortcomings for posterior fossa (PF): (i) deficient sensitivity (ii) subtle finding and
(iii) radiation exposure. Consequently, PF ischemic stroke lesions are missed at the
early stage which increasing the mortality rates. Nowadays, the development of
Computer-Aided Diagnosis (CAD) is increasingly becoming an important area in
stroke detection. Despite the rapid development of CAD in stroke diagnosis, no studies
have been found on stroke detection in PF. Until today, manual delineation of ischemic
stroke in PF on NECT demands dealing with a large amount of data, which leads to
late prognosis. As the amount of image data generated by NECT is massive, Deep
Learning (DL) solutions are among the effective ways to deal with complex and large
amount of cross-sectional data. Therefore, a new diagnostic algorithm based on DL is
proposed for ischemic stroke detection in PF. The algorithm framework consists of
hybrid of improved Xception model and YOLO V2 detector to classify the PF slices
with ischemic and localise the infarction in classified slices, respectively. Following
that, a CAD system is established by integrating the proposed algorithmic framework.
The performance and effectiveness of the proposed algorithmic are evaluated by the
comparison with the gold standard provided by the radiologists. The proposed
algorithmic framework has shown to be less prone to overfitting and simultaneously
improves the detection performance than the original DL model. The results
demonstrate that the performance measure of 90.77% has been recorded for detection
rate with average processing time of 1.02 to 1.04 seconds per image. The developed
algorithm is reported to be reliable to assist the radiologist in ischemic PF diagnosis
which is important for future healthcare needs. |
|---|