Loss functions for imbalanced medical segmentation

WHAT THIS TESTS: It probes whether you understand the gradient dynamics of segmentation losses under extreme foreground-background imbalance, which is a daily reality in medical imaging where a lesion is a fraction of a percent of the scan.

A GOOD ANSWER COVERS: Plain cross-entropy averages loss over every pixel, so when a tumor is one percent of the image the loss is dominated by easy background pixels and the network can reach very low loss while barely segmenting the tumor at all. Dice Loss equals one minus the Dice coefficient computed over predicted and true masks; because it is a ratio of overlap it is largely insensitive to the absolute size of the region and pushes directly toward the overlap metric you report. Focal Loss multiplies cross-entropy by a modulating factor that shrinks the contribution of confidently correct pixels, forcing learning to concentrate on the hard, rare foreground. A common practical choice is a compound loss such as Dice plus weighted cross-entropy.

COMMON WRONG ANSWERS: Saying just add class weights and stopping there. Claiming Dice is always strictly better, ignoring that its gradients can be unstable for tiny or empty masks. Confusing Focal Loss with simple hard-negative mining rather than a continuous down-weighting of easy examples.

LIKELY FOLLOW-UPS: How does the Focal gamma parameter change behavior. Why combine Dice with cross-entropy instead of using Dice alone. How do you handle slices that contain no tumor at all. What does Tversky loss add when you want to trade precision against recall.

ONE CONCRETE EXAMPLE: For a tumor occupying two hundred of twenty thousand pixels, a network predicting all background gets near-zero cross-entropy yet a Dice score of zero, so cross-entropy gives almost no useful gradient. Switching to Dice plus weighted cross-entropy makes the loss fall only when overlap genuinely improves, so the model actually learns the lesion boundary instead of the trivial all-background solution.