Training data: 500 images/masks over 100 epochs.
Accuracy calculation was done by comparing the number of pixels between the ground truth and the predicted mask (ignoring the background).
Accuracy: Achieved 96.7% accuracy.
Experimented using Dice loss function compared with the default categorical cross-entropy loss function to see if accuracy or F1-score would increase with unsuccessful results.
I used F1-score (the balance between precision and recall) as a heuristic measurement to compare models.
Image example:
Mask example:
Output:
