MRI image processing (2)
Computer-aided detection (CAD) is a tool to help radiologists detect and diagnose diseases. Currently, radiologists have to manually find the abnormalities. There are 620 images for each patient, which is time-consuming and the risk of misdiagnosis is high due to human error. For patients with Gliob...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2020
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/144637 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Computer-aided detection (CAD) is a tool to help radiologists detect and diagnose diseases. Currently, radiologists have to manually find the abnormalities. There are 620 images for each patient, which is time-consuming and the risk of misdiagnosis is high due to human error. For patients with Glioblastoma Multiforme, it is crucial to have fast and accurate detection to start treatment plans as early as possible. Previous studies have been conducted to detect abnormalities automatically by using machine learning techniques or visual saliency methodology. However, the models were either too complicated or they did not use ground truth to evaluate their accuracy or they were working at a 2D level.
The proposed algorithm (PR) uses visual saliency methodology to detect abnormalities in MRI images.
The algorithm first uses three MRI sequences T1c, T2 and FLAIR to generate a pseudo-colored image channel (RGB) which is then converted to CIE L*a*b color space. The produced volumetric image is divided into cubes of size 4 x 4 x 4, 8 x 8 x 8 and 16 x 16 x 16. Each cube is represented by its mean intensity. Color differences and spatial differences between each cube pair are taken into consideration to produce a 3D saliency map. All three planes (xy, yz, xz) are also taken into consideration.
The produced 3D saliency map is evaluated by comparing it with ground truth and 2D model. It was established that the proposed algorithm performs better than its 2D model. |
---|