Glossary

This glossary collects recurring terms used across the book. Architecture pages still explain terms in context, but this page gives readers one place to check definitions.

Term	Meaning
Semantic segmentation	Segmentation where each pixel or voxel receives a class label, without separating different objects of the same class.
Instance segmentation	Segmentation where separate objects receive separate IDs, not just foreground/background labels.
Promptable segmentation	Segmentation where user guidance such as points, boxes, masks, text, or memory controls which mask the model returns.
Foundation model	A broadly pretrained model adapted to downstream segmentation workflows; in this book, the term does not imply clinical readiness.
Volumetric segmentation	Segmentation over 3D images or patches, usually represented as (B, C, D, H, W), with through-plane context.
Dense prediction	A task where the model produces a prediction at each pixel, voxel, or spatial location instead of one label for the whole image.
Tensor shape notation	The project convention for naming tensor dimensions, such as (B, C, H, W) for 2D batches and (B, C, D, H, W) for 3D batches.
Logits	Raw model scores before sigmoid or softmax. Use logits directly with logits-aware losses.
Probability map	A logit map converted to values interpretable as probabilities.
Feature map	An internal tensor of learned channels at a particular spatial resolution.
Voxel	A volume element in 3D data, analogous to a pixel in 2D data.
Encoder-decoder	A model pattern where an encoder compresses spatial features and a decoder restores dense spatial predictions.
Downsampling / upsampling	Downsampling reduces spatial resolution to increase context; upsampling restores resolution for dense output.
Transposed convolution	A learned upsampling operation often used in decoders to increase spatial resolution.
Skip connection	A path that carries earlier features into a later layer, often to restore spatial detail.
Bottleneck	The deepest, lowest-resolution representation in an encoder-decoder model.
Output stride	The ratio between input resolution and an internal feature-map resolution.
ASPP	Atrous Spatial Pyramid Pooling, a multi-scale context module using dilated convolutions.
Atrous / dilated convolution	A convolution that spaces out kernel samples to increase receptive field without adding pooling.
Residual block / shortcut	A block that adds a shortcut path around layers so the block can learn a correction to its input.
Attention gate	A module that uses learned weights to filter or emphasize features, often before merging skip features into a decoder.
Deep supervision	Extra training losses attached to intermediate decoder outputs to guide earlier layers during optimization.
Tokenization and patch embedding	The process of turning image or volume patches into token features consumed by Transformer-style layers.
Shifted-window attention	Swin-style local attention where window boundaries shift between blocks so neighboring windows can exchange context.
Prompt	User-provided guidance such as a point, box, mask, text label, or memory state.
Prompt-conditioned mask decoder	A decoder that combines image features with prompt features to produce the requested mask.
Memory-conditioned prompting	Prompting that uses stored context from prior images, slices, frames, or interactions to influence the next mask.
Patch-based training	Training on cropped image or volume regions instead of full scans to control memory use and sampling.
Sliding-window inference	Applying a model to overlapping patches and stitching the patch predictions into a full-size output.
Voxel spacing / resampling	Voxel spacing is the physical size represented by each voxel; resampling changes an image or mask to a chosen spacing or grid.
Dataset fingerprint	A dataset summary, such as shapes, spacing, modalities, labels, and intensity behavior, used to plan preprocessing or model settings.
Dataset leakage	Evaluation contamination caused by related patients, cases, slices, patches, preprocessing statistics, or duplicates crossing data splits.
External validation	Evaluation on data from a separate site, scanner, institution, acquisition protocol, or domain beyond internal splits.
Threshold policy	The chosen rule for converting probabilities into binary predictions; changing it changes precision and sensitivity tradeoffs.
Dice score	An overlap metric: twice the intersection divided by predicted plus target area.
Soft Dice loss	A differentiable overlap loss computed from probabilities rather than hard masks, commonly used for imbalanced segmentation.
IoU	Intersection over Union, also called Jaccard.
Sensitivity	The fraction of true foreground captured by a prediction, also called recall.
Specificity	The fraction of true background correctly rejected by a prediction.
HD95	The 95th percentile Hausdorff distance, a boundary-distance metric less sensitive to a single worst outlier than full Hausdorff distance.
Surface Dice	A boundary agreement metric that counts surfaces as matching when they are within a chosen distance tolerance.
Average precision	An instance-segmentation metric that summarizes precision and recall over one or more IoU thresholds.
Sparse annotation / loss mask	Training with only some labeled pixels or voxels, where unlabeled locations must be ignored or handled explicitly in the loss.
PHI / DICOM header	Protected health information and medical-image metadata that may contain patient identifiers or clinical details.
Clinical-readiness claim	A claim that a model is safe, approved, or ready for diagnosis or deployment; this project treats such claims as out of scope.
Synthetic data	Artificially generated examples used for tests, demos, or tutorials instead of real patient, clinical, or private data.
Star-convex polyhedron	A 3D StarDist-style object shape represented by distances from an object center along fixed rays.
Radial distance map	A map of center-to-boundary distances used to parameterize star-convex object proposals.
Non-maximum suppression	A postprocessing step that removes overlapping duplicate object proposals while keeping the strongest candidates.
Flow field	A dense vector output that indicates how pixels or voxels should move or group during instance reconstruction.
Gradient tracking	Following a predicted vector or gradient field from pixels or voxels toward object centers or regions.
Normalized cut	A graph-partitioning objective that groups similar neighboring pixels or voxels into coherent regions while penalizing weak separation.