Instance Dice Score in Segmentation

• Instance Dice score is a statistical measure that quantifies overlap between predicted and ground-truth segments, providing values from 0 (no overlap) to 1 (perfect overlap).
• It is defined as twice the intersection over the sum of prediction and ground-truth cardinalities, with averaging schemes like macro and micro applied to handle multiple instances.
• Differentiable surrogates such as soft-Dice loss improve deep learning models by optimizing segmentation performance, particularly in imbalanced and small-instance settings.

The instance Dice score, or Dice similarity coefficient (DSC), is a set-based statistical measure quantifying agreement between predicted and ground-truth segmentation instances, particularly in medical imaging. Defined as twice the intersection cardinality divided by the total cardinalities of prediction and ground-truth, it provides a normalized overlap metric ranging from 0 (no overlap) to 1 (perfect agreement). The instance Dice score serves as a de facto standard for segmentation quality assessment and has driven development of loss surrogates in deep learning-based segmentation frameworks (Bertels et al., 2019).

1. Formal Definition and Averaging Schemes

Given a single object in an image, let $A \subset \Omega$ denote the set of pixels (or voxels) for the ground-truth instance, and $B \subset \Omega$ for the predicted instance. The Dice coefficient is:

$DSC(A,B) = \frac{2|A \cap B|}{|A| + |B|}$

This formulation enforces values in $[0,1]$ to assess similarity. For images with multiple instances, instance-level Dice can be computed as:

• Macro-average: Averaging $DSC(A_k, B_k)$ over each matched ground-truth instance $k$, yielding equal weight per object.
• Micro-average: Pooling all voxels from all instances and computing a global Dice, thus weighting larger instances more heavily.

Unmatched objects (false positives or negatives) are assigned Dice of zero. Division by zero is conventionally resolved by defining $DSC=1$ when both $|A| + |B| = 0$ (both empty), and $DSC=0$ if only one is empty (Bertels et al., 2019).

2. Relationship to the Jaccard Index

The Dice coefficient is closely related to the Jaccard index (Intersection over Union, IoU), with a bijective mapping:

$$J(A,B) = \frac{|A \cap B|}{|A \cup B|}, \quad J(A,B) = \frac{DSC(A,B)}{2 - DSC(A,B)}, \quad DSC(A,B) = \frac{2J(A,B)}{1 + J(A,B)}$$

Risk minimization under one measure bounds the risk under the other, indicating that optimizing one typically provides strong control over the other (Bertels et al., 2019).

3. Differentiable Surrogates and Loss Optimization

Since $DSC(A,B)$ is discrete, direct optimization via gradient descent is not feasible in deep learning. The “soft-Dice” loss offers a differentiable surrogate for CNN training. Given predicted probabilities $p_i \in [0,1]$ and ground-truth labels $g_i \in \{0,1\}$:

$$\mathcal{L}_{\mathrm{soft\text{-}Dice}}(P,G) = 1 - \frac{2\sum_i p_i g_i}{\sum_i p_i + \sum_i g_i}$$

The partial derivative with respect to $p_j$ is:

$$\frac{\partial \mathcal{L}_{\mathrm{soft\text{-}Dice}}}{\partial p_j} = -2 \frac{g_j (S_p + S_g) - S_{pg}}{(S_p + S_g)^2}$$

where $S_p = \sum_i p_i$, $S_g = \sum_i g_i$, $S_{pg} = \sum_i p_i g_i$. For $g_j=1$, the gradient increases $p_j$ proportionally to the overlap deficit, while for $g_j=0$, the gradient penalizes false positives (Bertels et al., 2019).

4. Cross-Entropy versus Metric-Sensitive Surrogates

Weighted cross-entropy (CE) loss, often used to handle class imbalance, cannot emulate the instance Dice or Jaccard measures for all cases. Relative-approximation bounds demonstrate that no optimal choice of cross-entropy weighting provides bounded multiplicative or additive approximation to the Dice (or Jaccard) loss universally. The Dice and Jaccard measures relatively approximate each other (with worst-case relative error 1 and absolute error ≃0.17), but the approximation between Dice and any weighted Hamming similarity deteriorates as object size decreases. This result emphasizes that, for small instances, weighted cross-entropy may result in arbitrarily poor Dice performance (Bertels et al., 2019).

5. Empirical Evaluation and Practical Impact

Empirical results on five segmentation tasks, evaluated using UNet and DeepLab-type CNNs, used various loss functions:

• Classical cross-entropy (CE)
• Weighted cross-entropy (wCE)
• Soft-Dice loss (sDice)
• Soft-Jaccard loss (sJaccard)
• Lovász-softmax loss (convex surrogate for IoU)

Metric-sensitive losses (sDice, sJaccard, Lovász-softmax) consistently outperform CE and wCE on instance-level Dice metrics, particularly in imbalanced settings. There is no significant statistical difference among the metric-sensitive surrogates in final Dice performance. Analysis across object sizes confirms that metric-sensitive losses outperform CE-based losses, validating the theoretical risk minimization arguments (Bertels et al., 2019).

6. Computation and Reporting Guidelines

To ensure fair and reproducible evaluation, the following practices are advised:

• Object matching: Assign each ground-truth object $A_k$ to a prediction $B_k$ (commonly via maximal IoU).
• Averaging: Clearly state macro or micro averaging; for imbalanced distributions, stratify by object-size deciles.
• Empty objects: Use standard conventions for DSC when one or both sets are empty.
• Reporting: State the averaging scheme and report both mean ± standard deviation (or median/interquartile range).
• Normalization: Absolute Dice may depend on resolution, object complexity, and labeling protocol; fixing instances ensures cross-dataset consistency.

Adoption of metric-sensitive loss functions and adherence to rigorous reporting protocols are critical for robust segmentation evaluation in medical imaging research (Bertels et al., 2019).

7. Limitations and Theoretical Insights

Neither the Dice nor the Jaccard index can be exactly emulated by a weighted cross-entropy loss. All tested surrogates (soft-Dice, soft-Jaccard, Lovász-softmax) are equivalent up to a multiplicative factor, and no unique optimal surrogate emerges across tasks. These metrics are particularly sensitive to class imbalance and small object sizes, a key consideration for risk minimization and practical deployment (Bertels et al., 2019).