Diabetic Retinopathy Image Classification

Post on 15-Aug-2015

31 views 0 download

Tags:

Transcript of Diabetic Retinopathy Image Classification

Diabetic Retinopathy Image Classification

Dr. Yusheng Feng, Dr. Artyom Grigoryan, John Jenkinson

Research Objective• Produce a scheme for the detection of

exudates in optical fundus images for the automated diagnosis of diabetic retinopathy

Group Members• Dr. Yusheng Feng

• (Mechanical Engineering, SiViRT* Director)

• Dr. Artyom Grigoryan • (Electrical and Computer Engineering)

• John Jenkinson • (Electrical and Computer Engineering, SiViRT*)

• *SiViRT – Center for Visualization, Simulation, and Real-Time Prediction

Original Pathological Image

Figure 1. Original Pathological RGB Image from Messidor Database

Red Channel and Correlation Mask

This mask is correlated with each image to locate the Optic Disc

Figure 3. Correlation Mask

Figure 2. Red channel of original RGB Image

Correlation Image and Disk Mask

Figure 4. Correlation Image with brightest point where largest correlation occurs

Figure 5. Mask created by setting pixels within 150 radius from circle center to zero

Green Channel with Removed Disk

Figure 6. Green channel image with optic disc removed by mask

Top-hat Transform Filter

Figure 7. Top-hat transform working as an regional adaptive threshold on the Green channel image

Contrast Stretching Top-hat Image

Figure 8. Contrast stretched image to fill the dynamic range of the top-hat enhanced image, and example region of exudates displayed with high intensity enveloped in green

Figure 9. Region of contrast-stretched Image displayed exudate groups circled in red

Threshold Binary Image

Figure 10. Background removed by applying an image threshold to contrast-stretched image

Morphological Opening

Figure 11. Morphological opening with structuring element disk of size 1 for stage one noise removal (coarse objects)

Median Filtering

Figure 12. Median filter window size 7x7 for stage 2 noise removal (fine objects)

Border Pixel Removal

Figure 13,14. Red channel image (left) displayed with border region displayed in white. Border region is applied to Median Filtered image (right) and sets all pixels within the border to zero intensity

Kirsch’s Edge Analysis

Figure 15,16. Kirsch’s edge analysis (left) is used to develop a probability map for exudate candidates (right)

Exudate Candidate Map

Figure 17,18. The average value of an edge from Kirsch’s edge analysis is assigned to every pixel in that object (left) with a region of objects displayed (right)

Probability

Figure 19. The candidate map is then scaled so that the value of each object represents its probability of being an exudate. The features of each object extracted for classification (next slide) are weighted by the value of their exudate probability

Feature Extraction

• Density Feature: Ratio of Object Area (constant region) to Bounding Box (slashed region)

• Area, Mean Intensity, and Variance Features: The constant region represents an object. The features corresponding to this object representation are the Area: number of pixels in the object, Mean Intensity: average intensity of the Green channel pixels for the object and Variance: see Mean Intensity

• Axes Ratio Feature: The ratio of the Major axis, represented by the line AB to the Minor Axis represented by the line CD

Figure 20. Density Feature Figure 21. Area, Mean Intensity, Variance Figure 22. Axes Ratio

Classification• Trained: Decision Trees,

Support Vector Machines, Nearest Neighbor Classifiers, and Ensemble Classifiers.

• SVM with Quadratic Kernel performed the best with 0.7238 Area under the Receiver Operating Characteristic Curve with Hold-One-Out 0.2 Threshold Validation.

Next Steps• Improve classification performance by

adjusting feature selection and optimizing segmentation.

• Optimize segmentation for additional databases: Diaretdb1, HEI-MED, and e-optha EX

THANK YOUQuestions? Comments?