Download presentation
Presentation is loading. Please wait.
Published byBaby Marte Modified over 9 years ago
1
Kyle Marcolini MRI Scan Classification
2
Previous Research For EEN653, project devised based on custom built classifier for demented MRI brain scans Minimal processing methods implemented in preprocessing/segmentation stage Minimal features extracted based on image characteristics Classifier was ~80-90% accurate in determining demented brain scans
3
Proposal Using previously built classifier and scan database, implement methods for preprocessing and feature extraction Attempt to increase classification accuracy without changing the classifier Focus solely on processing of the scans
4
Brain Database Oasis brain database Each file contains brain scan and text file, which contains: ○ Info on scanned subject (age, sex) ○ MMSE score (cognitive impairment test) ○ CDR (rating of dementia) ○ Final categorization of (either none, slight, or full-on dementia)
5
Previous Preprocessing Methods Increase the brightness and contrast by a predetermined factor for all scans Set threshold levels to diminish 256 possible intensity levels to 5 This intruduced a lot of noise No further noise removal, average, or smoothing Resulting images varied in brightness and threshold level
6
Previous Features Extracted Image mean (average intensity value) Symmetry (healthy brains tend to be more symmetrical) Gradient mean and variance (edges of brain) Normalized black area in hippocamous (center black area, usually darker means more dementia)
7
Proposed Processing Methods Adaptive brightness and contrast based on prior scan’s color mean Averaging filter to remove unwanted noise pixels Deblurring for darker scanse with a lot of inherent noise and blur Better thresholding (still thinking of better way than select ranges of intensities to round to a single value)
8
Additional Features Incorporated some prior features applied to filtered scans, the gradients of the scans Frequency-based analysis rather than just spatial Incorporate wavelet transform-based features Detection techniques to search for abnormalities or potential lesions
9
Predicitions The k-nearest neighbor classifier from before successfully classified many of the noisy scans With filtered and consistent scans, I hope to achieve a program that is consistently greater than 95% accurate
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.