Computer-Aided Detection of Pulmonary Nodules in Computed Tomography

Computer-aided detection (CAD) of pulmonary nodules in computed tomography (CT) has transformed the early identification of potential lung malignancies and benign lesions. By augmenting radiological interpretation with algorithmic analysis, CAD aims to improve sensitivity for small or subtle nodules while managing false-positive rates to avoid unnecessary follow-up. Modern systems harness machine learning, particularly deep convolutional neural networks, to distinguish true nodular structures from surrounding anatomy, to segment lesions in three dimensions and to characterise growth over time. Performance is typically evaluated by per-scan and per-nodule metrics such as area under the receiver operating characteristic curve, nodule-level sensitivity at fixed false positives per scan, and free-response ROC analysis. Clinical integration of CAD addresses global shortages of experienced thoracic radiologists and standardises screening in high-risk populations, with applications in lung cancer screening programmes, incidental nodule detection in routine CT, and workflow optimisation through concurrent or second-reader paradigms. Despite demonstrable gains in sensitivity, challenges remain in reducing false positives, validating systems across diverse scanners and demographics, and defining cost-effective deployment strategies. Ongoing research seeks to refine algorithms for subsolid and vessel-attached nodules, to incorporate radiomic features for malignancy prediction, and to embed CAD seamlessly into picture archiving and communication systems.

Research from Nature Portfolio

Recent studies have demonstrated a deep learning-based AI system capable of detecting both benign and malignant pulmonary nodules in clinically indicated chest CT scans outside organised screening programmes. The system achieved sensitivities exceeding 94% for benign nodules and 97% for primary lung cancers at approximately one false positive per scan, matching or surpassing expert radiologist performance. External validation across multiple centres underscored its robustness to scanner variations, while comparison with a panel of thoracic radiologists confirmed its potential as an adjunct tool for incidental nodule assessment and early lung cancer detection.

Computer-Aided Detection of Pulmonary Nodules in Computed Tomography publication trend

The graph below shows the total number of articles in computer-aided detection of pulmonary nodules in computed tomography across all publications each year (not limited to Nature Index journals).

Technical terms

Pulmonary nodule: A small, round or oval lesion in the lung parenchyma, typically less than 30 mm in diameter.

Convolutional neural network (CNN): A class of deep learning model well suited to image analysis, using hierarchical filters to extract features.

Free-response receiver operating characteristic (FROC) curve: A plot of sensitivity versus average false positives per image, used to assess detection systems.

Segmentation: The process of delineating the boundaries of a nodule within imaging data, often in three dimensions.

False-positive rate: The frequency at which non-nodule structures are incorrectly identified as nodules by the CAD system.

Area under the ROC curve (AUROC): A scalar measure of diagnostic discrimination, representing the probability that a system ranks a positive case higher than a negative one.

References

1. Deep learning for the detection of benign and malignant pulmonary nodules in non-screening chest CT scans. Communications Medicine (2023).
2. Pricing and cost-saving potential for deep-learning computer-aided lung nodule detection software in CT lung cancer screening. Insights into Imaging (2023).
3. External validation of the performance of commercially available deep-learning-based lung nodule detection on low-dose CT images for lung cancer screening in Japan. Japanese Journal of Radiology (2024).
4. A Comprehensive Review of Performance Metrics for Computer-Aided Detection Systems. Bioengineering (2024).
5. Performance of computer-aided detection of pulmonary nodules in low-dose CT: comparison with double reading by nodule volume. European Radiology (2012).
6. The impact of trained radiographers as concurrent readers on performance and reading time of experienced radiologists in the UK Lung Cancer Screening (UKLS) trial. European Radiology (2017).
7. Validation of a deep learning computer aided system for CT based lung nodule detection, classification, and growth rate estimation in a routine clinical population. PLOS ONE (2022).