Choosing a Segmentation Metric When the Tumour Is 1% of Pixels

A U-Net segments tumours in 2D slices. On a 100,000-pixel slice the ground-truth tumour covers 1,000 pixels (1%). The model predicts 800 tumour pixels, of which 600 overlap the ground truth. A colleague reports "99.4% accuracy" and wants to ship; another proposes training with plain pixel-wise cross-entropy and tracking mIoU.

1. Fill the confusion counts and compute pixel accuracy, tumour IoU, tumour Dice, background IoU and mIoU. Compare with a model that predicts all background.
2. Show the relation between IoU and Dice with these numbers, and explain how they behave differently when averaged over a test set.
3. Choose a metric (or set of metrics) and a training loss for this problem and justify them. What do you do with slices that contain no tumour?
4. Explain why the U-Net's skip connections matter for this task.