Dual Glossary

The same word can mean different things in clinical practice and machine learning. This glossary pairs a concise definition with the distinction most likely to affect an imaging AI project.

Clinical and imaging terms for engineers

Term	Meaning and practical distinction
Screening	Assessment of a defined population without a specific presenting symptom; prevalence and follow-up differ from diagnostic cohorts.
Diagnostic examination	Imaging directed at a clinical question, symptom, or prior finding; not interchangeable with screening.
Indication	The clinical reason for an examination, or the specified authorized use of a device; identify which meaning is intended.
Finding	An observation such as an opacity or lesion; it is not necessarily a diagnosis.
Differential diagnosis	Plausible explanations for findings in context, not a list of independently established diseases.
Reference standard	The method used to establish the evaluation target; it may combine expert review, pathology, follow-up, or other evidence.
Incidental finding	An observation outside the original examination question; it still needs an appropriate review and follow-up pathway.
Triage	Prioritization of attention or workflow; a negative triage result does not automatically rule out disease.
Recall	A request for additional assessment after screening; not synonymous with a cancer diagnosis.
Lesion	A localized abnormality; its definition and boundaries depend on the task.
Longitudinal	Across time for the same person or target; requires correct identity, timing, and comparability.
Acquisition protocol	The settings and sequence of measurements used to obtain an examination.
Series	A grouping of DICOM objects; one study may contain many series with different roles.
Reconstruction	Producing an image from measured signals; distinct from generating a report or classifying disease.
Windowing	Mapping an intensity range to display brightness; generally not a change to the underlying measurement.
Voxel	A sample in a volumetric grid; its physical dimensions may differ along each axis.
Registration	Estimating spatial correspondence between images; registration error can masquerade as biological change.
Partial-volume effect	Mixing of contributions within limited spatial resolution, affecting boundaries and measurements.
Contrast phase	Timing of acquisition relative to contrast administration; different phases are different model inputs.
Hounsfield unit	CT attenuation scale relative to water; not a universal scale for all tomographic imaging.
SUV	A normalized PET uptake measure with a specified convention, units, and timing assumptions.
Ejection fraction	Fraction of end-diastolic ventricular volume ejected; the arithmetic depends on valid source measurements.
Whole-slide image	A digitized tissue section, often a multiresolution pyramid; it is not the complete specimen or case.
Grading	Categorizing severity or morphology under a defined system; integer labels may not be comparable across systems.
Ungradable	Inadequate for the specified assessment; it must not be relabeled as normal.
PACS / RIS / LIS	Picture archive, radiology information system, and laboratory information system; their identities and workflow roles must be integrated.

Machine-learning terms for clinicians

Term	Meaning and practical distinction
Training / validation / test	Data used to fit a model, select its settings, and estimate held-out performance; repeated tuning on the test set undermines independence.
Data leakage	Information unavailable at the intended decision point enters development or evaluation, inflating apparent performance.
Domain shift	A change in the input or clinical setting between development and use, such as a different scanner or population.
Distribution drift	A change over deployment time; it is a signal for investigation, not automatically proof of clinical harm.
Classification	Assigning a category or score to an input; it does not necessarily localize the evidence.
Detection	Identifying candidate locations or events; a box is not a precise boundary or histological diagnosis.
Segmentation	Assigning labels to pixels or voxels; labels require anatomical and coordinate definitions.
Regression	Predicting a numerical quantity; correlation with a reference does not guarantee agreement.
Embedding	A learned numerical representation; dimensions are not usually directly interpretable clinical measurements.
Foundation model	A broadly pretrained model adapted to downstream tasks; breadth of pretraining is not validation for a specific clinical use.
Fine-tuning	Updating model parameters for a new task or dataset; it can improve specialization and introduce new failure modes.
Checkpoint	A saved parameter state; preprocessing, labels, and dependencies are still needed to reproduce a result.
Calibration	Agreement between predicted probabilities and observed frequencies in a defined setting.
Discrimination	Ability to rank or separate outcomes; good discrimination does not imply good calibration.
Operating point	The selected threshold or policy at which sensitivity, specificity, and workload are evaluated.
Sensitivity / specificity	Positive detection among reference positives and negative detection among reference negatives; specify the unit and population.
Positive predictive value	Fraction of positive outputs that are reference positive; depends on prevalence and operating point.
AUROC	Ranking performance across thresholds; it does not select a clinically acceptable threshold.
AUPRC	Area under the precision–recall curve; its interpretation depends strongly on positive prevalence.
Dice score	Overlap measure for masks; a high average can conceal missed small lesions or important boundary errors.
Uncertainty	Lack of confidence due to data, labels, or model limitations; a score is useful only if its interpretation is validated.
Abstention	An explicit refusal to produce a supported result; requires a defined downstream action.
Weak supervision	Training labels are indirect or coarse, such as report-mined labels or slide-level labels for tiles.
Multiple-instance learning	Learning from groups of instances with a group-level label, common in whole-slide analysis.
Explainability	Methods intended to help understand outputs; a heatmap is not automatically a causal explanation.
Hallucination	Generated content unsupported by the available evidence; fluent wording does not make it factual.
Agent	A system that selects and sequences tools or actions; permissions and stopping conditions are part of its safety.
Latency / throughput	Delay for one task versus tasks processed per unit time; optimizing one may worsen the other.
Idempotency	Repeating the same request does not create unintended duplicate effects.
Provenance	The traceable relationship among source data, transformations, model versions, outputs, and review.
Rollback	Returning to a previously controlled version or workflow after a problematic change.

Words to clarify in every project

Accuracy may refer to a classification metric, measurement agreement, or general correctness. Name the actual endpoint. Validation may mean a development split, technical verification, external clinical evaluation, or an institutional acceptance process. State which one was performed. Approved may mean internal permission or a particular regulatory pathway; do not use it as a generic synonym for model availability. Normal may mean no target finding in an adequate examination, not the absence of every disease.