Alzheimer's disease classification using Deep learning Neural a Network and Gamma Correction
- 1. Classification of Alzheimer’s disease subjects from MRI using Deep Learning Neural Network and Gamma Correction Yubraj Gupta ID = 20177743 Master in Information and Technology Engineering Digital Media Computing Lab Prof. Goo-Rak Kwon
- 2. Content • Dataset • Dataset preparation • Data Augmentation • Proposed Method • Image pre-processing • Label and it’s features
- 3. Dataset • Oasis dataset which consist 100 Alzheimer’s Disease patients record and 316 Non- Alzheimer’s Disease patients records • ADNI dataset was performed by 1.5T and 3T scanner. • Here for my research I am using 1.5T which consist of 453 AD and 748 Non-AD patient record. Normal image Alzheimer’s disease Fig.1: Normal and AD MRI brain images
- 4. Dataset preparation • As, I already mention in previous slide that there are 100 data for AD and 316 for Non- AD patients. • So to increase the dataset’s : I use the following steps • Step 1: In this step I just extract three view of original AD and NON-AD images. • Step 2: And, In second step I follow the Data augmentation method which is a kind of data increasing method. • Step 3: This step is a Pre-processing steps and in this case I compare gamma correction with adaptive histogram (CLAHE) technique and found that for medical image gamma correction provide better image visibility. Axial view Coronal viewSagittal view Original image view’s with their slices
- 5. Data Augmentation Image at 0 degree 0 degree flip images
- 6. Continue… • Image at 90 degree 90 degree flip images
- 7. Continue… • Image at 180 degree 180 degree flip images
- 8. Continue… • Image at 270 degree 180 degree flip images
- 9. Dataset sizes • So, now total dataset will be • For OASIS : AD – 143200 images Non-AD – 452512 images • For ADNI : AD – 645072 images Non-AD – 1065072 images
- 10. Gamma correction vs CLAHE Original image CLAHE processed image Gamma Correction processed image Similarity index shows comparison between original and CLAHE processed image Similarity index shows comparison between original and Gamma correction processed image 1. AND, we can see that gamma correction technique perform better than CLAHE
- 11. Proposed Method 3D MRI image 60 slice Non-AD AD Fig: Proposed Blocked Diagram Data Augmentation Technique at 0,90,180,270 degree 0-256 Axial-slice 0-166 Sagittal-slice 0-256 Coronal-slice Image pre-Processing Normalization ALEXNET
- 12. Training Accuracy at 15 epoch of OASIS dataset Epoch Iteration Mini_batch Loss Mini-batch Accuracy Base Learning Rate 15 93600 0.0603 96.88% 0.00004 15 93650 0.0239 100.00% 0.00004 15 93700 0.0559 98.44% 0.00004 15 93750 0.095 95.31% 0.00004 15 93800 0.0466 98.44% 0.00004 15 93850 0.0854 95.31% 0.00004 15 93900 0.529 96.88% 0.00004 15 93950 0.045 98.44% 0.00004 Total there is 20 Epoch for OASIS Dataset with 64 Mini-batch size Figure shows training progress at 15th Epoch Training Accuracy plot’s
- 13. Training Accuracy at 4 epoch of ADNI dataset Epoch Iteration Mini_batch Loss Mini-batch Accuracy Base Learning Rate 4 65600 0.2519 85.94% 0.0001 4 65650 0.2757 93.75% 0.0001 4 65700 0.1615 95.31% 0.0001 4 65750 0.1643 90.63% 0.0001 4 65800 0.1067 98.44% 0.0001 4 65850 0.1905 90.63% 0.0001 4 65900 0.1453 93.75% 0.0001 Total there is 20 Epoch for ADNI Dataset with 64 Mini-batch size Figure shows training progress at 4th Epoch Training Accuracy plot’s
- 14. Result • Oasis dataset result has not come yet, it’s in processing stage • Adni dataset result has not come yet, it’s in processing stage
- 15. Thank You