Glossary¶
53 terms used throughout the curriculum, in plain clinical language. Glossary entries link to the chapter where the term is introduced. Source-of-truth Word document at glossary/CNN_Med_Imaging_Glossary_2026.docx.
CNN & Medical Imaging Glossary: 2026 Edition
Clinical AI 2026: Computer Vision for Radiologists
Activation Function: A mathematical function applied to the output of a neuron to introduce non-linearity and determine its activation level. Common activation functions include the sigmoid, ReLU, and tanh functions.
Artificial Neural Network (ANN): A computational model inspired by the human brain, consisting of interconnected nodes (neurons) organized in layers to process and learn from data.
Attention Map: A visualization of the attention weights in a transformer model, showing which image patches the model focuses on when making a prediction. Attention maps provide an intrinsic form of explainability for Vision Transformers - unlike GradCAM, which is applied post-hoc, attention weights are a natural byproduct of the model's computation.
Attention Mechanism: The ability of a model to focus on specific parts of the input when producing output. In vision, this means learning which image patches are relevant to each other for a given task, allowing the model to 'pay attention' to distant but related regions simultaneously - much like how an experienced radiologist's eye naturally jumps between related findings across an image.
AUC (Area Under the Curve): AUC is a common evaluation metric used in binary classification tasks to assess the performance of a model's predictive ability. It measures the overall quality of the model's ranking or classification predictions. AUC represents the area under the Receiver Operating Characteristic (ROC) curve, which plots the true positive rate (sensitivity) against the false positive rate (1 - specificity) at various classification thresholds. AUC ranges from 0 to 1, where a higher value indicates better performance. An AUC of 0.5 suggests random performance, while an AUC of 1 represents a perfect classifier. The AUC also helps determine the optimal operating threshold for clinical use: a screening tool might prioritize high sensitivity (catching all cases), while a confirmatory test might prioritize high specificity (minimizing false alarms).
Backpropagation: An algorithm that answers the question: 'Which weights contributed most to the error, and how should we adjust them?' An algorithm used to train neural networks by propagating the error from the output layer back to the input layer, adjusting the weights and biases of the network to minimize the error.
Batch Normalization: A technique used to normalize the inputs of a neural network layer by adjusting and scaling the activations. It improves the stability and convergence of the network during training.
Binary Cross Entropy (BCE) loss: A loss function used in binary classification that measures the difference between predicted probabilities and true labels. The formula is: BCE = -[y × log(p) + (1-y) × log(1-p)], where y is the true label (0 or 1) and p is the predicted probability. BCE penalizes confident wrong predictions heavily - predicting 0.99 when the true label is 0 incurs a much larger loss than predicting 0.6. During training, the model adjusts its weights to minimize the average BCE across all training examples.
Calibration: The degree to which a model's predicted probabilities match actual observed frequencies. A well-calibrated model that predicts 80% probability of malignancy should be correct approximately 80% of the time. Poor calibration - where a model is overconfident or underconfident - can lead to inappropriate clinical decisions, such as unnecessary biopsies or missed diagnoses.
Classes: Classes are a fundamental concept in object-oriented programming. A class is a blueprint or template that defines the attributes (data) and behaviors (methods) of objects. It encapsulates data and the operations that can be performed on that data. Objects created from classes are instances of those classes.
Confidence Interval: A range of values, computed from the data, expected to contain the true value of a metric (e.g., AUC) with a specified probability - typically 95%. A narrow confidence interval (e.g., AUC 0.94-0.96) indicates a precise estimate; a wide one (e.g., 0.85-0.99) suggests uncertainty, often due to small sample size. Reporting confidence intervals is essential for clinical validation and regulatory submissions.
Contrastive Learning: A self-supervised training technique that teaches a model to produce similar representations for augmented versions of the same image and different representations for different images. The model learns meaningful visual features without any labeled data by distinguishing 'similar' from 'dissimilar' pairs - enabling pretraining on vast amounts of unlabeled medical images.
ConvNeXt: A modernized CNN architecture that incorporates design principles from Vision Transformers (such as larger kernel sizes and layer normalization) while retaining the convolutional structure. ConvNeXt demonstrates that CNNs can match transformer performance when updated with modern training techniques, offering a practical middle ground between traditional CNNs and ViTs.
Convolutional Neural Network (CNN): A type of neural network designed for analyzing visual data, such as images. CNNs apply convolutional operations to extract features hierarchically and are widely used in tasks like image classification and object detection. While CNNs dominated medical imaging from 2015-2022, Vision Transformers and hybrid architectures have become increasingly competitive, especially when large pretraining datasets are available.
Data Leakage: When information from the test or validation set inadvertently influences model training, leading to overly optimistic performance estimates that do not reflect real-world performance. Common sources include splitting data after applying transforms, having images from the same patient in both training and test sets, or using future data to predict past events.
Data transformations: An important step in preparing the data for deep learning because it helps improve the performance and effectiveness of the models. It can involve a wide range of techniques, including but not limited to the following: Preprocessing: This involves standardizing the data by applying normalization, scaling, or centering techniques to ensure that the features have similar ranges or distributions. Preprocessing may also involve handling missing values, removing outliers, or encoding categorical variables. Feature extraction: This process involves extracting meaningful and informative features from the raw data. It can be achieved using techniques such as dimensionality reduction (e.g., Principal Component Analysis), feature selection (e.g., selecting the most relevant features), or feature engineering (creating new features based on domain knowledge). Augmentation: Data augmentation is commonly used in computer vision tasks, where new training samples are generated by applying various transformations to the existing data. This can include operations such as flipping, rotation, scaling, cropping, or adding noise to the images, resulting in an augmented dataset that improves the model's ability to generalize and handle variations in the input data.
Deep Learning: A subset of machine learning that utilizes artificial neural networks with multiple hidden layers to learn and extract complex patterns and features from data.
DICOM (Digital Imaging and Communications in Medicine): The international standard file format and communication protocol for medical images. DICOM files store both pixel data (the image itself, often at 12-16 bit depth) and metadata (patient demographics, acquisition parameters, study information). Understanding DICOM is essential for building clinical AI pipelines that ingest images from hospital PACS systems.
Distribution Shift: A change in the statistical properties of data between the training environment and the deployment environment. Common causes include different scanner manufacturers, imaging protocols, patient demographics, or disease prevalence between institutions. Distribution shift is the primary cause of real-world performance degradation in clinical AI and is why external validation across multiple sites is essential.
Dropout: A regularization technique used in neural networks to prevent overfitting. During training, randomly selected neurons are ignored or "dropped out" with a certain probability, forcing the network to learn redundant representations and improving generalization.
Early Stopping: A regularization technique that halts training when performance on the validation set stops improving for a specified number of consecutive epochs (the 'patience' parameter). This prevents the model from memorizing the training data (overfitting) and typically results in better generalization to new, unseen patients. Most clinical AI training runs use early stopping rather than a fixed number of epochs.
EfficientNet: A family of CNN architectures that systematically scales network depth, width, and input resolution together using a compound scaling method discovered through neural architecture search. EfficientNets achieve strong accuracy while using fewer parameters and less computation than comparable models, making them particularly suitable for resource-constrained clinical environments.
Epoch: One complete pass through the entire training dataset during model training. In each epoch, the model sees every training example once (in randomized batches). Training for multiple epochs allows the model to iteratively refine its weights. Typical clinical AI training runs 50-200 epochs with early stopping, though tutorials often use 4-10 epochs for speed.
F1 Score: The F1 score is a measure of a model's accuracy in binary classification tasks, considering both precision and recall. It provides a balance between precision and recall by calculating the harmonic mean of these two metrics. The F1 score is particularly useful when the dataset is imbalanced, meaning that the number of instances from different classes is significantly different. The formula to calculate the F1 score is: F1 Score = 2 * ((Precision * Recall) / (Precision + Recall)) The F1 score ranges from 0 to 1, where a value of 1 indicates perfect precision and recall, while 0 represents the worst performance. The F1 score is a useful metric when the focus is on achieving a balance between minimizing false positives (precision) and false negatives (recall). It is commonly used in tasks where both precision and recall are important, such as information retrieval, document classification, and spam detection.
Feature Extraction: Using a pretrained model as a fixed feature extractor - the backbone weights are frozen and only a new classification head is trained on task-specific data. This is the fastest and cheapest approach to transfer learning, ideal when labeled clinical data is very limited (fewer than a few hundred images), as it avoids the risk of overfitting the backbone.
Fine-Tuning: Adapting a pretrained model to a new task by continuing training on task-specific data, typically with a smaller learning rate to avoid overwriting previously learned features. Fine-tuning can range from updating only the last few layers (partial fine-tuning) to updating the entire model (full fine-tuning), with the choice depending on how similar the target task is to the pretraining task.
Foundation Model: A large model pretrained on broad data (e.g., millions of medical images) using self-supervised learning, designed to be adapted to many downstream tasks with relatively small labeled datasets. In medical imaging, foundation models like BiomedCLIP, RAD-DINO, and CheXzero represent the current state-of-the-art, enabling strong performance on new classification, segmentation, or retrieval tasks with minimal task-specific training.
Functions: Functions are blocks of organized, reusable code that perform a specific task. They can take input parameters (arguments) and optionally return a value. Functions help in modularizing code, improving code readability, and promoting code reuse.
GradCAM (Gradient-weighted Class Activation Mapping): A technique that produces a heatmap highlighting which regions of an input image were most important for a model's prediction. It works by computing the gradients of the target class score flowing into the final convolutional layer and using them to weight the feature maps. GradCAM is widely used in medical imaging to provide visual explanations that radiologists can interpret, though it should be treated as hypothesis-generating rather than definitive proof of the model's reasoning.
Graphics Processing Unit (GPUs): It is a specialized chip that can accelerate the processing of computer graphics and visual data. GPUs have evolved to become powerful parallel processors capable of performing complex calculations and computations which speeds up the training of AI models.
Hyperparameters: Parameters of a deep learning model that are not learned from the data but are set by the user before training. Examples include learning rate, batch size, number of layers, and activation functions.
Learning Rate Scheduler: An algorithm that adjusts the learning rate during training according to a predefined strategy - for example, reducing it when validation loss plateaus (ReduceLROnPlateau), decaying it by a fixed factor each epoch (StepLR), or following a cosine annealing curve. Proper learning rate scheduling often improves both convergence speed and final model performance.
Loss Function: A function that measures the discrepancy between the predicted output of a model and the true target output. The goal of training a deep learning model is to minimize the loss function.
Modules: Modules are files or collections of code that encapsulate related functionality, variables, and classes. They provide a way to organize code into reusable units. Modules can be imported and used in other scripts or programs to access the contained functionality.
Multi-Head Attention: Running multiple attention operations in parallel, each learning different types of relationships between image patches. The outputs are concatenated and combined, allowing the model to capture diverse patterns simultaneously - for example, one head might focus on texture similarity while another captures spatial proximity between anatomically related structures.
Optimizer: An algorithm that adjusts the parameters (weights and biases) of a neural network based on the gradients of the loss function. Common optimizers include Stochastic Gradient Descent (SGD), Adam, and RMSprop.
Overfitting: A phenomenon where a machine learning model performs well on the training data but fails to generalize to unseen data. It occurs when the model captures noise or irrelevant patterns instead of the underlying patterns in the data.
Packages: In programming, a package is a collection of related modules or files that are organized together for easy management and reuse. Packages provide a way to structure and organize code in a hierarchical manner. They help in avoiding naming conflicts and provide a logical grouping of related functionality.
Patch Embedding: The process of dividing an image into non-overlapping patches (e.g., 16×16 pixels) and projecting each patch into a fixed-length vector representation. This converts a 2D image into a sequence of vectors that can be processed by a transformer - analogous to tokenizing a sentence into words before feeding it to a language model.
Positional Encoding: Information added to patch embeddings to preserve spatial position, since transformers have no inherent notion of order or location. Without positional encoding, the model would treat a patch from the upper-left corner of a chest X-ray identically to one from the lower-right, losing critical anatomical context.
Precision: Precision measures the proportion of correctly predicted positive instances (true positives) out of all instances predicted as positive (true positives + false positives). It focuses on the accuracy of the positive predictions. Precision = (True Positives) / (True Positives + False Positives) A high precision indicates that the model has a low rate of false positives, meaning it is good at correctly identifying positive instances.
Recall: Recall, also known as sensitivity or true positive rate, measures the proportion of correctly predicted positive instances (true positives) out of all actual positive instances (true positives + false negatives). It focuses on capturing as many positive instances as possible. Recall = (True Positives) / (True Positives + False Negatives) A high recall indicates that the model has a low rate of false negatives, meaning it is good at capturing the positive instances.
Saliency Map: A visualization that highlights pixels in the input image that most influence the model's output, computed by taking the gradient of the output with respect to the input pixels. Brighter pixels in the saliency map have more influence on the prediction. Saliency maps are simple to compute but can be noisy and less interpretable than GradCAM for clinical applications.
Scripts: Scripts refer to executable files containing a sequence of instructions or commands written in a programming language. They are typically used to automate tasks or perform specific operations. Scripts can be run directly by an interpreter or a scripting engine without the need for compilation.
Self-Attention: A mechanism where each element in a sequence (e.g., each image patch) computes a relevance score with every other element. This allows the model to dynamically weigh relationships - for instance, a patch showing a lung nodule can attend to the mediastinal region to assess lymph node involvement, regardless of the physical distance between them in the image.
Self-Supervised Learning: A training paradigm where the model learns visual representations from unlabeled data by solving pretext tasks - for example, predicting masked image regions, matching augmented views of the same image (contrastive learning), or predicting the relative position of image patches. This enables training on vast amounts of unlabeled medical data that would be prohibitively expensive to manually label.
Sensitivity and Specificity: Two complementary metrics for evaluating diagnostic performance. Sensitivity (true positive rate) measures the proportion of actual positives correctly identified - critical for screening tests where missing a disease is costly. Specificity (true negative rate) measures the proportion of actual negatives correctly identified - critical for confirmatory tests where false alarms lead to unnecessary procedures. The trade-off between sensitivity and specificity is visualized by the ROC curve.
Sessions: Sessions are a concept commonly used in interactive programming environments or frameworks. A session represents a particular instance or context in which code is executed. It can store and maintain the state of variables, objects, or other resources between multiple interactions or executions of code.
SHAP (SHapley Additive exPlanations): A game-theoretic approach to explaining individual predictions by computing the contribution of each feature (or image region) to the prediction. SHAP provides theoretically grounded, additive attribution values - the sum of all feature contributions equals the model's output. While mathematically rigorous, SHAP is computationally expensive for large images and is more commonly used for tabular data in clinical AI.
Subgroup Analysis: Evaluating model performance separately across different patient subgroups defined by demographic characteristics (age, sex, race), clinical variables (disease severity, comorbidities), or technical factors (scanner type, imaging protocol). Subgroup analysis reveals performance disparities that aggregate metrics can mask and is increasingly required by regulatory bodies such as the FDA for AI/ML medical device submissions.
Transfer Learning: A technique in which a pre-trained neural network, usually trained on a large dataset, is used as a starting point for a new task. The network's learned features can be transferred or fine-tuned for the new task, often resulting in improved performance, especially when the new dataset is small. The evolution of transfer learning: ImageNet pretraining (2015-2019) → self-supervised pretraining (2020-2022) → medical imaging foundation models (2022-present).
Variables: Variables are named containers used to store data in computer programs. They have a name, a value, and a specific data type. Variables can be assigned values that can be modified during program execution. They provide a way to store and manipulate data dynamically within the program's scope.
Vision Transformer (ViT): A neural network architecture that applies the transformer framework (originally designed for text processing) to image classification by dividing images into fixed-size patches and processing them as a sequence. ViTs have matched or exceeded CNN performance on many medical imaging benchmarks when pretrained on large datasets, though they typically require more data or stronger pretraining than CNNs to perform well.