Automatic Lung Nodule Annotation – Deep-Learning Vocabulary

A curated set of key vocabulary terms covering imaging basics, deep-learning architectures, attention modules, datasets, evaluation metrics, and workflow concepts for automatic lung-nodule annotation.

Computed Tomography (CT)

Cross-sectional X-ray imaging technique that creates detailed 3-D views of internal anatomy and is the primary modality for lung nodule evaluation.

Low-Dose CT (LDCT)

CT protocol that reduces radiation exposure and is widely used for lung-cancer screening programs.

Pulmonary Nodule

A small, roughly spherical opacity in the lung that may represent benign or malignant tissue and is often the first radiologic sign of lung cancer.

Detection (nodule task)

Automated identification of candidate nodule locations, typically outputting bounding boxes and confidence scores.

Segmentation (nodule task)

Pixel- or voxel-wise outlining of a nodule to obtain an exact mask and enable quantitative size and volume measurements.

Classification / Risk Stratification

Assigning detected nodules to categories such as benign vs. malignant or estimating future cancer risk.

Deep Learning (DL)

Subset of machine learning that uses multilayer neural networks to learn hierarchical feature representations directly from data.

Convolutional Neural Network (CNN)

Neural architecture that applies learnable filters over local regions of an image or volume, excelling at spatial feature extraction.

U-Net

Symmetric encoder–decoder CNN with skip connections, widely considered the standard baseline for medical-image segmentation.

Vision Transformer (ViT)

Model that splits images into patches and uses self-attention to capture global context, offering long-range dependency modeling.

DEtection TRansformer (DETR)

Anchor-free transformer framework that formulates object detection as a direct set-prediction problem without non-maximum suppression.

Attention Mechanism

Neural component that weights important spatial or channel features more heavily, helping the model ‘focus’ on relevant image regions.

Feature Pyramid Network (FPN)

Multi-scale architecture that merges high-resolution spatial details with low-resolution semantic features for robust object detection.

Multi-Scale Architecture

Network design that processes information at several resolutions to handle nodules ranging from a few millimetres to several centimetres.

Anchor-Free Detection

Detection strategy that predicts object centers or keypoints directly, eliminating predefined anchor boxes (e.g., SCPM-Net).

False Positive (FP)

Model output incorrectly labeled as a nodule; high FP rates burden radiologists and reduce clinical usability.

Dice Similarity Coefficient (DSC)

Overlap metric used in segmentation; calculated as 2 × |A∩B| / (|A|+|B|).

Competition Performance Metric (CPM)

Average sensitivity at seven fixed FP/scan rates, standard for evaluating lung-nodule detectors in LUNA16.

Free-response ROC (FROC)

Curve plotting sensitivity versus average false positives per scan across thresholds for detection assessment.

Area Under the Curve (AUC)

Integral of the ROC curve, measuring a classifier’s ability to distinguish classes across all thresholds.

Self-Supervised Learning (SSL)

Training paradigm that learns representations from unlabeled data using pretext tasks, later fine-tuned with limited annotations.

Federated Learning (FL)

Privacy-preserving technique where multiple institutions train a shared model by exchanging weight updates instead of raw data.

Models Genesis

3-D self-supervised pre-training framework that restores transformed patches to learn generic medical-image representations.

Semantic Genesis

Enhanced SSL approach combining self-discovery, self-classification, and self-restoration to embed anatomical semantics.

Active Learning

Human-in-the-loop strategy where the model selects the most informative unlabeled samples for expert annotation.

Semi-Automated Annotation

Workflow in which preliminary AI labels are edited or confirmed by radiologists, reducing manual effort.

Computer-Aided Detection (CADe)

System that highlights suspicious regions for the radiologist to review, emphasizing sensitivity over specificity.

Computer-Aided Diagnosis (CADx)

Tool providing quantitative measurements or malignancy probabilities to support clinical decision making.

Ground-Glass Opacity (GGO)

Subtle, hazy lung lesion with partial transparency that poses detection and segmentation challenges.

Squeeze-and-Excitation (SE) Block

Channel-attention module that adaptively recalibrates feature-map importance via global pooling and gating.

Spatial Attention

Mechanism that emphasizes informative spatial locations within a feature map (e.g., CBAM spatial branch).

Channel Attention

Mechanism that weights entire feature channels according to their relevance for the task (e.g., SE blocks).

Coordinate Attention (CA)

Attention module encoding long-range dependencies along height and width while preserving positional information.

Transformer-CNN Hybrid

Architecture combining CNN encoders for local detail with transformer modules for global context (e.g., MCAT-Net).

Sphere Representation

Anchor-free modeling choice in SCPM-Net that uses center and radius to depict 3-D nodules instead of bounding boxes.

Lung-DETR

Deformable DETR variant tailored to sparse pulmonary-nodule detection using Maximum-Intensity-Projection preprocessing.

Swin-Tempo

Method treating CT as video slices; uses Swin Transformer plus GRUs for temporal lung-nodule detection.

CPLOYO

YOLO-based detector with CAMF and MSCAF modules for advanced multi-scale feature fusion in nodule finding.

Nested 3-D FCN

Dense-skip CNN segmentation architecture (Kido et al.) designed to better capture complex or attached nodules.

Multi-Threshold Feature Separation (MFS)

Input pre-processing that splits CT voxels into multiple HU ranges to enhance edge and density cues for segmentation.

Electronic Health Record (EHR) Integration

Using structured or free-text clinical data alongside images to improve false-positive reduction or risk modeling.

Prior CT Integration

Exploiting earlier scans to capture temporal changes in nodules, boosting malignancy-risk estimation.

Uncertainty Quantification

Model output of confidence or predictive variance to inform clinicians where AI predictions may be less reliable.

nnU-Net

Self-configuring U-Net framework that automatically adapts preprocessing, architecture, and training to a given dataset.

Sensitivity (Recall)

Proportion of actual nodules correctly detected or segmented by the model.

Precision

Proportion of positive detections that are true nodules, complementing sensitivity in evaluation.

mean Average Precision (mAP)

Detection score averaging precision over recall levels and IoU thresholds, common in object-detection benchmarks.

LIDC-IDRI

Public CT dataset of 1,018 scans with multi-reader nodule annotations, widely used for segmentation and classification.

LUNA16

Challenge subset of LIDC-IDRI (888 scans) providing standardized labels for nodule detection benchmarking.

NLST

National Lung Screening Trial dataset containing longitudinal LDCT scans and clinical outcomes for risk-model validation.

NELSON

Large European lung-cancer screening trial dataset offering diverse imaging protocols and population characteristics.

Dataset Bias

Systematic representation imbalance that can lead to unfair or non-generalizable model performance.

Explainable AI (XAI)

Methods that provide human-understandable reasoning or visualizations for model predictions to foster trust.

Human-in-the-Loop

Collaborative paradigm where AI assists but final judgment remains with human experts, enhancing safety and acceptance.

Edge-Device Deployment

Running trained models on resource-constrained hardware near data sources, necessitating model compression and optimization.

Data Augmentation

Artificial expansion of training data via transformations (e.g., rotation, noise) to improve model robustness.

Domain Shift

Performance drop when a model trained on one imaging domain (scanner, protocol, population) is applied to another.

Anchor

Predefined box template used in many detectors to hypothesize object locations and sizes.

Region Proposal Network (RPN)

CNN sub-module that generates candidate bounding boxes for object detectors such as Faster R-CNN.

Bounding Box

Rectangular (or cuboidal) region predicted around a nodule candidate during detection.

Segmentation Mask

Binary or multi-class image/volume where target pixels or voxels are labeled 1 and background 0.

3-D CT Volume

Stack of axial CT slices providing volumetric representation of the thorax for nodule analysis.

Hounsfield Unit (HU)

Quantitative scale for CT attenuation values; useful for tissue characterization and threshold-based processing.