Segmentation of Cornea Cells using U-Net Architecture

• Segmented the microscopic images of corneal endothelium cells and labeled each pixel as cell interior, cell border, or background.
• Developed and trained a 32x32 image patch-segmenting 24-layer U-Net model to an accuracy of 80% in training and 72% in validation.
• Reconstructed test segments by applying a sliding window operation on 500x500 test images while employing a 32x32 U-Net model.
• Plotted the ROC Curves and got an area under the curve (AUC) of 0.869 with the training set patches and 0.839 with the testing set patches.

NOTE: Open Final_Project.ipynb to see the full training, testing & evaluation processes.

The Network Architecture:

Segmentation & Reconstruction Quality:

Segmented 32x32 Patches output from the network:

ROC Curves:

Training:

Testing: