Residency · Residency · Diagnostic Radiology
Artificial Intelligence in Radiology: Foundations and Current Applications
Introduction
Artificial intelligence (AI) is transforming the practice of radiology by augmenting image interpretation, streamlining workflow, and enabling quantitative analysis at a scale impossible for the human reader alone. Understanding the foundational concepts of AI, machine learning, and deep learning is now a core competency for radiology residents. This lecture provides a framework for understanding how these technologies work, where they are currently deployed, and what their realistic capabilities and limitations are.
Foundational Concepts
Artificial Intelligence, Machine Learning, and Deep Learning
Artificial Intelligence (AI) is a broad field encompassing any computer system that performs tasks normally requiring human intelligence. Machine Learning (ML) is a subset of AI in which algorithms learn patterns from data without explicit programming, with performance improving as more training data is provided. Deep Learning (DL) is a subset of ML that uses artificial neural networks with multiple layers (hence "deep") to automatically learn hierarchical feature representations from raw data.
Supervised vs. Unsupervised Learning
Supervised learning trains models on labeled data (for example, chest X-rays labeled as "pneumothorax present" or "absent") and is the most common paradigm in radiology AI. Unsupervised learning identifies patterns in unlabeled data, such as through clustering or anomaly detection. Semi-supervised and self-supervised learning leverage small amounts of labeled data combined with large amounts of unlabeled data.
Convolutional Neural Networks (CNNs)
CNNs are the dominant architecture for medical image analysis. Convolutional layers apply learned filters to detect features (edges, textures, shapes) at increasing levels of abstraction. Pooling layers reduce spatial dimensions. Fully connected layers produce the final classification or regression output. CNNs can perform classification (such as benign vs. malignant), detection (such as locating a nodule), and segmentation (such as delineating tumor boundaries).
Transformer Models and Foundation Models
Vision transformers (ViTs) apply attention mechanisms to image patches, capturing global context. Foundation models (large models pretrained on massive datasets) can be fine-tuned for specific radiology tasks. Multimodal models integrate imaging data with clinical text (reports, EHR data).
Training, Validation, and Testing
The training set (typically 60-70% of data) is used to optimize model parameters. The validation set (typically 10-20%) is used to tune hyperparameters and prevent overfitting. The test set is held-out data used to evaluate final model performance and must be independent of training. External validation on data from a different institution or population is essential for assessing generalizability. Overfitting occurs when a model performs well on training data but poorly on new data and is mitigated by regularization, data augmentation, and dropout.
Performance Metrics
| Metric | Definition | Use |
|---|---|---|
| Sensitivity (Recall) | True positive rate | Ability to detect disease |
| Specificity | True negative rate | Ability to exclude disease |
| AUC | Area under ROC curve (0.5-1.0) | Overall discriminative ability |
| PPV | TP / (TP + FP) | Prevalence-dependent positive prediction |
| NPV | TN / (TN + FN) | Prevalence-dependent negative prediction |
| F1 Score | Harmonic mean of precision and recall | Useful for imbalanced datasets |
| Dice coefficient | 2x overlap / total | Segmentation task evaluation |
Sensitivity (recall) is the true positive rate and measures the ability to detect disease. Specificity is the true negative rate and measures the ability to exclude disease. Area Under the ROC Curve (AUC) measures overall discriminative ability, where 1.0 is perfect and 0.5 is chance. Positive Predictive Value (PPV) and Negative Predictive Value (NPV) depend on disease prevalence. The F1 Score is the harmonic mean of precision and recall and is useful for imbalanced datasets. The Dice coefficient is an overlap metric for segmentation tasks.
Current FDA-Cleared Applications in Radiology
Chest Imaging
Applications include pneumothorax detection on chest radiographs (prioritizing the worklist for urgent reads), pulmonary nodule detection and characterization on CT (computer-aided detection and lung cancer screening with Lung-RADS integration), and tuberculosis screening on chest X-ray (deployed in resource-limited settings).
Neuroimaging
Large vessel occlusion (LVO) detection on CT angiography alerts the stroke team for thrombectomy candidates and is among the most impactful clinical AI applications. Intracranial hemorrhage detection on non-contrast head CT provides triage and worklist prioritization. Quantitative brain volumetry automates hippocampal and regional brain volume measurements for dementia evaluation.
Breast Imaging
Mammography AI includes detection of suspicious lesions, density assessment, and risk prediction. Studies show AI can achieve radiologist-level sensitivity and may reduce false positives.
Musculoskeletal
Fracture detection on radiographs (wrist, hip, spine) reduces missed fractures, particularly in emergency settings. Bone age assessment is automated from hand radiographs.
Cardiac
Applications include coronary artery calcium scoring from non-gated chest CTs and CT fractional flow reserve (CT-FFR), an AI-derived functional assessment of coronary stenosis.
Workflow Integration
Worklist prioritization uses AI to flag critical findings (pneumothorax, LVO, hemorrhage) and move them to the top of the radiologist's queue. Quantitative reporting integrates automated measurements (tumor volume, ejection fraction, brain volumes) into structured reports. Pre-population of reports generates AI-derived preliminary findings for radiologist review and editing. Quality assurance provides a second-reader function to catch potential misses before report finalization.
Limitations and Challenges
Dataset bias means models trained on one population may perform poorly on others (different scanners, patient demographics, disease prevalence). Lack of explainability reflects the "black box" nature of deep learning models; Grad-CAM and saliency maps provide some interpretability but are imperfect. Automation bias refers to radiologists potentially over-relying on AI output, reducing their own vigilance. The regulatory pathway requires FDA 510(k) or De Novo clearance, with post-market surveillance for performance drift. Integration challenges involve compatibility with PACS, RIS, and EHR systems and compliance with DICOM and HL7 FHIR standards.
Key Clinical Pearls
AI in radiology is a decision-support tool, not a replacement for the radiologist; the radiologist remains responsible for the final interpretation. External validation on diverse, multi-institutional datasets is the most important indicator of a model's real-world reliability. The highest-impact AI applications currently are those that triage urgent findings (LVO, pneumothorax, hemorrhage) by reducing time to diagnosis. Always consider the prevalence of disease in your practice when interpreting AI performance metrics because high AUC does not guarantee high PPV in low-prevalence settings. Radiologists should be actively involved in AI development, validation, and governance to ensure clinical relevance and patient safety.
References
- Hosny A, Parmar C, Quackenbush J, et al. Artificial Intelligence in Radiology. Nat Rev Cancer. 2018;18(8):500-510.
- Rajpurkar P, Chen E, Banerjee O, et al. AI in Health and Medicine. Nat Med. 2022;28(1):31-38.
- LeCun Y, Bengio Y, Hinton G. Deep Learning. Nature. 2015;521(7553):436-444.
- ACR Data Science Institute. AI Central. Available at: https://aicentral.acrdsi.org.