Oncology AI Glossary

This glossary defines the artificial-intelligence and machine-learning terms that appear across OncoWiz, written for clinicians and trainees rather than for engineers. Each entry opens with a one-sentence definition that stands on its own, followed by what the term actually means in oncology practice and where it is most often misread. Terms are listed alphabetically and each one has its own link.

Algorithmic bias

Algorithmic bias is a systematic error that makes a model perform worse for some groups of people than for others. In oncology it usually enters through the training data rather than the mathematics: a model taught on screening images from one population, one scanner vendor or one hospital’s case mix inherits whatever that sample over- or under-represents, and reports its confident answers in the same tone for the patients it was never really taught about.

Area under the ROC curve (AUC-ROC)

The area under the ROC curve is a single number between 0.5 and 1.0 summarising how well a model separates two classes across every possible decision threshold. It is the most quoted metric in oncology AI papers and the most over-read: an AUC of 0.95 says the model ranks cases well, but says nothing about whether its probabilities are calibrated, how it behaves at the one threshold you would actually deploy, or how it performs in a population where the disease is far rarer.

Artificial intelligence (AI)

Artificial intelligence is the broad field of building computer systems that perform tasks normally requiring human intelligence, such as recognising patterns in images, interpreting language or making predictions. In clinical oncology the term almost always refers to machine learning in practice, and it is worth reading it as an umbrella rather than a single technology, because the evidence needed to trust a segmentation tool differs entirely from what is needed to trust a language model.

Attention mechanism

An attention mechanism is a component of a neural network that lets the model weigh which parts of its input matter most for each part of its output. It is the core idea behind transformers, and in pathology it is what allows a model to reason over a whole slide by focusing on a handful of diagnostically informative tiles; attention maps are often shown as explanations, though what a model attends to is not necessarily what drove its decision.

Automation bias

Automation bias is the human tendency to accept a computer’s recommendation over one’s own judgement, including when the recommendation is wrong. It is the dominant safety risk in clinical AI deployment rather than a footnote to it: a decision-support tool that is right 95% of the time can still cause net harm if clinicians stop scrutinising the 5%, which is why prospective evaluation measures the clinician-plus-model system and not the model alone.

Black box

A black box is a model whose internal reasoning cannot be inspected in terms a human can follow, even when its inputs and outputs are fully visible. Most high-performing deep learning models in oncology are black boxes in this sense, and the practical question is not whether the internals are legible but whether the model has been validated well enough, on populations close enough to yours, that you can justify acting on its output.

Calibration

Calibration is the agreement between a model’s stated probabilities and observed outcomes: among patients a well-calibrated model assigns a 20% risk, roughly 20% should experience the event. Discrimination and calibration are independent, and a model can rank patients beautifully while being systematically overconfident, which matters enormously when the output feeds a threshold-based clinical decision rather than a ranked worklist.

Class imbalance

Class imbalance is the situation where one outcome is far rarer than the other in the training data, as malignancy is in a screening cohort. It is the reason accuracy is a meaningless headline metric in oncology: a model that calls every screening mammogram normal is over 99% accurate and clinically worthless, so imbalanced problems are judged on sensitivity, positive predictive value and precision-recall curves instead.

Clinical decision support system (CDSS)

A clinical decision support system is software that presents clinicians with patient-specific assessments or recommendations to inform a decision, without making that decision itself. The distinction matters for regulation as much as for practice: systems that let a competent clinician independently review the basis of a recommendation are often regulated more lightly than those whose output cannot practically be checked.

Computational pathology

Computational pathology is the analysis of digitised tissue slides by machine learning, covering tasks from tumour detection and grading to predicting molecular status directly from morphology. It depends on whole-slide imaging, and its most striking results, such as inferring mutation status from a haematoxylin and eosin slide, are also the ones that generalise least reliably across laboratories with different staining and scanning practices.

Computer-aided detection (CAD)

Computer-aided detection is software that marks regions of an image a radiologist should examine, leaving interpretation to the reader. It predates the deep learning era by decades, and the history of traditional mammography CAD, widely deployed before evidence showed it did not improve outcomes and may have increased recall rates, is the cautionary case study every oncology AI deployment should be measured against.

Confusion matrix

A confusion matrix is the table of true positives, false positives, true negatives and false negatives produced by a model at one decision threshold. Nearly every classification metric is derived from it, and reading the raw counts rather than the summary statistics is usually the fastest way to see whether a headline number is hiding a clinically unacceptable error pattern.

Convolutional neural network (CNN)

A convolutional neural network is a neural network architecture built to process images by learning small local filters and combining them into progressively more abstract features. CNNs underpin most published oncology imaging models, in radiology and pathology alike, and while transformer-based vision models increasingly match or beat them, the CNN remains the default baseline any new imaging result is expected to beat.

Data drift

Data drift is the gradual divergence between the data a deployed model receives and the data it was trained on, caused by new scanners, changed protocols, shifting referral patterns or a changed population. It degrades performance silently, without an error message, which is why a deployed clinical model needs ongoing monitoring rather than a single validation at go-live.

Deep learning

Deep learning is machine learning using neural networks with many layers, which learn useful representations directly from raw data rather than from features an expert specified in advance. It is responsible for nearly all recent progress in medical imaging AI, and its defining trade-off is that removing the need for hand-crafted features also removes the ability to inspect what those features are.

Dice coefficient

The Dice coefficient is a measure of overlap between a predicted region and a reference region, running from 0 for no overlap to 1 for exact agreement. It is the standard metric for tumour segmentation, and its main limitation is that it is dominated by large structures, so a model can post an excellent Dice score while systematically missing small lesions.

Digital biomarker

A digital biomarker is a measurable indicator of a biological or clinical state derived computationally from digital data such as imaging, sensor streams or electronic records. Radiomic signatures predicting treatment response are the oncology archetype; like any biomarker, one requires analytical validity, clinical validity and demonstrated clinical utility before it means anything, and most published candidates never clear the third.

Explainable AI (XAI)

Explainable AI is the set of techniques for making a model’s behaviour interpretable to humans, including saliency maps, attention visualisations and attribution methods. Explanations are best treated as hypotheses about the model rather than accounts of its reasoning: saliency maps for the same prediction differ substantially between methods, and a plausible-looking heatmap has repeatedly been shown to accompany a model relying on an artefact.

External validation

External validation is the evaluation of a model on data from institutions, populations or equipment entirely separate from those used in development. It is the single most informative thing a clinical AI paper can report and the most frequently omitted; performance typically drops on external data, and the size of that drop is a far better guide to real-world behaviour than the internal test result.

F1 score

The F1 score is the harmonic mean of precision and recall, condensing both into one number between 0 and 1. It is useful for comparing models on imbalanced tasks, but weights precision and recall equally, which rarely matches oncology, where the cost of missing a cancer and the cost of a false alarm are not remotely comparable.

False negative

A false negative is a case the model calls negative that is in fact positive, such as a malignancy the model reports as benign. In oncology screening and triage this is usually the error that causes direct patient harm, and it is the error a single accuracy figure most effectively conceals.

False positive

A false positive is a case the model calls positive that is in fact negative. In an oncology setting the resulting harm is real but different in kind from a missed cancer: unnecessary biopsies, imaging, cost and patient anxiety, all of which scale with the number of people screened.

Federated learning

Federated learning is a method of training one model across several institutions by exchanging model updates rather than patient data, so records never leave the site that holds them. It addresses a genuine governance obstacle to multi-centre oncology models, but does not by itself solve consent, model-inversion risk, or the practical difficulty of harmonising labels and protocols across sites.

Few-shot learning

Few-shot learning is a model’s ability to perform a new task from a handful of examples rather than a full training set. It matters in oncology because the conditions where help is most needed, rare tumours and rare presentations, are exactly those with too few cases to train on conventionally.

Foundation model

A foundation model is a large model pre-trained on a broad corpus of data and then adapted to many downstream tasks. Pathology and radiology foundation models trained on millions of unlabelled images are an active research front; their promise is strong performance from far fewer labels, and their open question is whether validation done on one downstream task tells you anything about safety on another.

Generalisation

Generalisation is a model’s ability to perform on data it has never seen, as opposed to reproducing what it memorised. It is the entire point of building a model and the property most easily faked, because any evaluation contaminated by overlap between training and test data, including two scans of the same patient split across both, reports generalisation that does not exist.

Generative AI

Generative AI is the class of models that produce new content such as text, images or audio rather than classifying or scoring existing content. In oncology its most credible near-term uses are administrative, including drafting documentation and summarising records, and its central failure mode is that a fluent, confident output carries no signal about whether it is correct.

Ground truth

Ground truth is the reference label a model is trained and judged against, such as a histopathological diagnosis, a radiologist’s annotation or a documented outcome. It is a floor on how good a model can honestly be measured to be: where the reference is itself an expert opinion with meaningful inter-observer variability, a model cannot be shown to exceed the reliability of the label it was taught.

Hallucination

A hallucination is fluent, confident output from a generative model that is factually wrong, including invented citations, dosages or findings. It is not a bug to be patched but a property of models trained to produce plausible continuations, which is why any clinical use of generative text requires the output to be verifiable against a source rather than trusted on its presentation.

Inference

Inference is the act of running a trained model on new data to obtain a prediction, as distinct from training the model in the first place. The distinction is practical: training is expensive and occasional, inference is cheap and continuous, and it is inference that has to meet a clinical workflow’s latency, reliability and audit requirements.

Interobserver variability

Interobserver variability is the extent to which qualified experts disagree with one another on the same case. It sets the realistic ceiling for AI evaluation in oncology, because a model judged against one reader’s annotations is being measured against a moving target, and reported superiority to a single expert may say more about that expert than about the model.

Large language model (LLM)

A large language model is a neural network trained on very large text corpora to predict the next token, which yields a general capacity to interpret and produce language. In oncology, LLMs are being applied to summarisation, coding, trial matching and patient communication, and their outputs need to be evaluated as text produced by a statistical process rather than as answers from a source of clinical knowledge.

Machine learning

Machine learning is the branch of artificial intelligence in which systems learn patterns from data instead of following rules written by a programmer. It is the technology behind almost everything currently called clinical AI, and its defining consequence is that the behaviour of the system is determined by the data it was shown, which makes data provenance a clinical safety question.

Model card

A model card is a short structured document stating a model’s intended use, training data, evaluation results, known limitations and performance across subgroups. It is the closest equivalent AI has to a prescribing information leaflet, and the absence of one for a tool being proposed for clinical use is itself informative.

Multimodal model

A multimodal model is one that takes in more than one type of data at once, such as imaging together with pathology, genomics and clinical notes. Oncology is a natural application because staging and treatment decisions are already multimodal, and the practical difficulty is that a model can lean almost entirely on the easiest modality while appearing to integrate all of them.

Natural language processing (NLP)

Natural language processing is the computational analysis of human language, used in oncology to extract structured facts from radiology reports, pathology reports and clinical notes. It is how the unstructured majority of the oncology record becomes analysable, and its errors are systematic rather than random, with negation and uncertainty language the classic sources of failure.

Negative predictive value (NPV)

Negative predictive value is the probability that a case the model calls negative really is negative. Unlike sensitivity and specificity it depends on prevalence, so an NPV measured in an enriched research cohort will not transfer to a screening population, where it is usually the number a clinician actually wants.

Neural network

A neural network is a model made of layers of simple interconnected units whose connection strengths are adjusted during training. The biological analogy that gives it its name is loose enough to be misleading; what matters clinically is that a network is a very flexible function fitted to data, and flexibility is exactly what allows it to fit the wrong thing.

Overfitting

Overfitting is what happens when a model learns the noise and idiosyncrasies of its training data instead of the underlying signal, performing excellently on that data and poorly on anything new. It is the most common reason an impressive published result fails to reproduce, and the standard defence, a strictly held-out test set, only works if it was genuinely held out.

Positive predictive value (PPV)

Positive predictive value is the probability that a case the model flags as positive really is positive. It depends heavily on prevalence, which is why a model with excellent sensitivity and specificity can still generate mostly false alarms when deployed in a low-prevalence screening setting.

Precision

Precision, in machine learning, is the proportion of positive predictions that are correct, and is identical to positive predictive value. The overlap in vocabulary between machine learning and clinical epidemiology causes regular confusion, and precision in this sense is unrelated to precision oncology.

Precision oncology

Precision oncology is the practice of tailoring cancer treatment to the molecular and clinical characteristics of an individual tumour and patient. It is a clinical paradigm rather than a computational one, but it generates the high-dimensional, multimodal data that machine learning is used to interpret, which is why the two terms so often appear together.

Prompt engineering

Prompt engineering is the practice of designing the input given to a generative model in order to get more reliable output. It is a real skill with real effects, and also a fragile one: a prompt tuned on one model version can behave differently on the next, which makes undocumented prompts a poor foundation for anything clinical.

Radiomics

Radiomics is the extraction of large numbers of quantitative features from medical images, such as texture, shape and intensity statistics, for use as predictive or prognostic markers. The field’s central methodological problem is that many features are unstable across scanners and acquisition protocols, so a signature that is not tested for reproducibility is usually measuring the scanner as much as the tumour.

Retrieval-augmented generation (RAG)

Retrieval-augmented generation is an architecture in which a language model retrieves relevant documents and generates its answer from them, rather than from its training weights alone. It is the most practical current mitigation for hallucination in clinical settings because it makes answers traceable to a source, though it fails in its own ways when retrieval returns the wrong document or the model ignores what it retrieved.

Segmentation

Segmentation is the task of delineating a structure in an image voxel by voxel or pixel by pixel, such as outlining a tumour and organs at risk. It is the most clinically embedded use of AI in oncology today, particularly in radiotherapy planning, where automatically generated contours are edited by a clinician rather than accepted outright.

Sensitivity (recall)

Sensitivity is the proportion of truly positive cases the model correctly identifies, and is the same quantity machine learning calls recall. In cancer detection it is usually the metric with the highest clinical stakes, because the cases it counts are the ones missed.

Software as a Medical Device (SaMD)

Software as a Medical Device is software intended for a medical purpose that performs that purpose without being part of a hardware medical device. It is the regulatory category most clinical AI falls into, and the deciding question for any given tool is its intended use as stated by the manufacturer, not the sophistication of the underlying model.

Specificity

Specificity is the proportion of truly negative cases the model correctly identifies. It trades off against sensitivity as the decision threshold moves, and choosing where to sit on that trade-off is a clinical and economic judgement about the relative cost of missed disease and false alarms, not a technical one.

Supervised learning

Supervised learning is training a model on examples that are paired with known correct answers, such as images labelled with a confirmed diagnosis. It accounts for most clinical AI, and its practical constraint is that expert labels in oncology are slow, expensive and variable, which is what makes label-efficient approaches attractive.

Training, validation and test sets

Training, validation and test sets are the three disjoint partitions of data used to fit a model, tune its settings, and finally measure it once. The discipline only holds if the split is made at the patient level; splitting by image or by slide lets the same patient appear on both sides and inflates the reported result.

Transfer learning

Transfer learning is reusing a model trained on one task as the starting point for another, typically by fine-tuning it on a smaller dataset. It is what makes medical imaging models trainable on the thousands of examples a hospital can realistically assemble rather than the millions the original pre-training required.

Transformer

A transformer is a neural network architecture built around attention, now dominant in language processing and increasingly strong in vision. Both large language models and most current pathology and radiology foundation models are transformers, so the term describes the shared machinery beneath tools that otherwise appear unrelated.

Unsupervised learning

Unsupervised learning is training a model on data with no labels, so that it discovers structure such as clusters or representations on its own. In oncology it is used to find subgroups in molecular or imaging data without presupposing what they are, and its results require careful validation because clustering algorithms will return clusters whether or not any exist.

Whole-slide imaging (WSI)

Whole-slide imaging is the digitisation of an entire glass pathology slide into a high-resolution image that can be viewed and analysed on screen. It is the precondition for computational pathology, and the resulting files are large enough, often gigapixels per slide, that storage and network capacity become genuine deployment constraints.

Zero-shot learning

Zero-shot learning is a model performing a task it was never explicitly trained on, guided only by a description of what is wanted. Large models often show real zero-shot ability, and in a clinical context that ability is precisely what has not been validated, so plausible zero-shot performance is a reason for evaluation rather than a substitute for it.