Focal Loss for Segmentation

Focal loss multiplies cross-entropy by a factor that fades easy background pixels to near zero, leaving rare edge and lesion pixels in charge.

Focal loss turns down the volume on confident background pixels so a few uncertain edge pixels dominate the gradient.

Why Does This Exist?

In dense masks $90$ percent or more of pixels are easy background the model masters in epoch one, yet plain cross-entropy keeps summing their small losses until they drown the few lesion and boundary pixels. Dice fixes this with overlap scoring; focal loss (Lin et al., 2017, adapted from detection to masks) fixes it by reweighting: multiply each pixel's cross-entropy by $(1 - p_t)^\gamma$ with $\gamma \approx 2$ , where $p_t$ is the predicted probability of the true class. Easy pixels fade; hard ones rule.

For the glossary entry see the focal loss definition. This page is the segmentation variant with per-pixel behaviour and numbers.

Think of It Like This

A coach who stops drilling mastered scales

A piano coach hears forty easy scales played right and two hard passages fumbled. Equal attention wastes the hour on scales. Focal loss hands the coach a volume knob wired to mastery: confident passages fade to a whisper, fumbled bars play loudly. Practice time flows to the failures.

Where it stops: the knob needs tuning. Too high a gamma mutes useful reinforcement and the easy passages drift back out of tune.

How It Actually Works

Per pixel with true-class probability $p_t$ and balance factor $\alpha_t$ :

$\mathcal{L}_{focal} = -\alpha_t (1 - p_t)^\gamma \log(p_t)$

With $\gamma = 2$ : a background pixel at $p_t = 0.95$ contributes $(0.05)^2 = 0.0025$ of its cross-entropy, while an edge pixel at $p_t = 0.3$ keeps $(0.7)^2 = 0.49$ . The easy pixel is downweighted roughly $200\times$ relative to the hard one before $\alpha$ balancing. $\alpha = 0.25$ for foreground is the standard starting point from RetinaNet practice.

Worked example

Two pixels, $\gamma = 2$ , no $\alpha$ . Easy: $p_t = 0.9$ , cross-entropy $-\log(0.9) \approx 0.1054$ , focal $0.01 \times 0.1054 \approx 0.0011$ . Hard: $p_t = 0.3$ , cross-entropy $\approx 1.2040$ , focal $0.49 \times 1.2040 \approx 0.5900$ . Ratio hard-to-easy rises from about $11\times$ to about $536\times$ . That is why boundaries finally move.

Code

1import math2
3def focal(pt: float, gamma: float = 2.0) -> float:4    return -((1 - pt) ** gamma) * math.log(pt)5
6print(round(focal(0.9), 4), round(focal(0.3), 4))7# -> (0.0011, 0.59)

Watch Out For

Gamma so high the model forgets easy regions

$\gamma = 5$ mutes background so hard that late-training drift reintroduces holes inside confident areas. Symptom: speckled interiors while edges look great. Fix: start at $\gamma = 2$ , and combine with Dice if interiors destabilize.

Applying detection alpha blindly to masks

RetinaNet's $\alpha = 0.25$ assumes box imbalance, not your organ ratio. Symptom: foreground recall stuck low despite focal loss. Fix: set $\alpha$ from your pixel frequencies and validate per-class recall, not just mean loss.

The Quick Version

Focal loss multiplies cross-entropy by $(1 - p_t)^\gamma$ per pixel.
Confident background pixels fade by $100\times$ or more; hard pixels dominate.
Standard start: $\gamma = 2$ , foreground $\alpha = 0.25$ , then tune to your frequencies.
Complements Dice: focal mines hard pixels, Dice guards global overlap.
Too much gamma destabilizes confident interiors.