Segmentation of Optical Disk and Cup in fundus

Segmentation of Optical Disk and Cup in fundus image using FCN Prithviraj Dhar Supervised by : Dr. Amit Kale Bosch Research and Technology Center, India

Dataset The original BOSCH dataset consists of 73 images We’ve used Training : 50, Validation : 10, Testing : 13 All images are slightly degraded using Gaussian blur, motion blur, disk blur and distortion. Such degradation does not alter the annotated boundaries In the final dataset we have Training : 300, Validation : 60, Testing: 78 (six times what we had initially)

Dataset Input Image GT (0 -cup, 1 -OD, 2 -Background)

Method The raw images are fed as input, and ground truth as labels A fully convolutional network (derived from VGG-16) is trained using the aforementioned inputs and labels The trained network is then used for ternary pixel classification Post processing : In the final output, Connected Components are calculated and only the component having maximum area is considered.

Network output

Network output (post - processing)

Boundary Error Calculation Boundaries are computed using canny edge detector For a fixed angle [0 to 360 degree], the distance D between the pixel on the GT and that on the predicted curve is computed This is done separately for OD and Cup For an image, the boundary error is the mean of D computed at every angle Here, the mean boundary error for cup is 5. 16 and that of OD is 0. 78

Boundary (Output samples) OD Error: 0. 78, Cup error: 5. 16 White- GT, Black - Predicted

Boundary (Output samples) OD Error: 0. 78, Cup error: 5. 16 White- GT, Black - Predicted

Future Work Evaluate FCN models after adding extra weights for the annulus (OD) region Evaluate two-stream FCN models Evaluate the importance of vessel-ness features for prediction of cup regions

Thank You