Dice Score (F1 Score)

The Jaccard Index is a foundational spatial overlap metric that quantifies the ratio of the intersection area to the union area between a predicted mask and ground truth. While the Dice Score gives twice the weight to the intersection, IoU provides a stricter, more conservative measure of overlap. The two are monotonically related: Dice = 2 × IoU / (1 + IoU). IoU is the primary evaluation standard for object detection and instance segmentation tasks, and it serves as the basis for the Average Precision (AP) metric in challenges like COCO.

The Tversky Loss is a generalization of the Dice Loss that introduces tunable weighting parameters α and β to control the penalty for false positives and false negatives independently. This is critical in medical imaging where the cost of missing a lesion (false negative) far outweighs over-segmentation. By adjusting the α and β hyperparameters, engineers can directly encode clinical risk tolerance into the optimization objective, making it a preferred loss function for lesion segmentation and tumor boundary delineation.

The Hausdorff Distance is a boundary-based metric that complements the Dice Score by measuring the maximum distance from any point on the predicted boundary to the nearest point on the ground truth boundary. While Dice quantifies volumetric overlap, Hausdorff captures the worst-case segmentation error—a critical safety metric in applications like Organ-at-Risk (OAR) segmentation for radiotherapy, where a single boundary excursion can damage healthy tissue. The 95th percentile Hausdorff Distance (HD95) is commonly reported to exclude outlier sensitivity.

The U-Net is the foundational convolutional neural network architecture for biomedical image segmentation, featuring a symmetric encoder-decoder structure with skip connections that concatenate high-resolution features from the contracting path to the expanding path. This design preserves fine spatial detail lost during downsampling, enabling precise boundary localization. Originally developed for neuronal structure segmentation in electron microscopy, U-Net has become the de facto standard and backbone for countless medical imaging variants including 3D U-Net, Attention U-Net, and nnU-Net.

nnU-Net is a self-configuring segmentation framework that automatically adapts preprocessing, network topology, training strategy, and post-processing to any given medical dataset without manual tuning. It systematically analyzes dataset fingerprints—including voxel spacing, intensity distribution, and class imbalance—to generate a bespoke pipeline. nnU-Net has dominated numerous medical segmentation challenges, demonstrating that robust, automated configuration often outperforms hand-crafted architectural innovations. It serves as a critical baseline for any new segmentation research.

Connected Component Analysis is a post-processing algorithm applied to binary segmentation masks to identify and label isolated contiguous regions. In medical imaging, it is routinely used to remove small spurious predictions—such as scattered false-positive voxels—by filtering components below a volume threshold. It also enables the separation of touching objects in instance segmentation tasks. Combined with the Dice Score during evaluation, it ensures that clinically irrelevant noise does not inflate perceived model performance.