Deep Learning Methods in Lung CT Image Segmentation

Deep learning has transformed the segmentation of lung structures in computed tomography by enabling fully automated delineation of lung parenchyma, airways and lobes with high accuracy and efficiency. Convolutional neural networks form the core of this transformation, with specialised architectures such as U-Net and its variants delivering precise boundary localisation through encoder–decoder pathways and skip-connections. Recent advances incorporate multi-resolution analysis, attention mechanisms and ensemble strategies to handle variations in scanner protocols, anatomical anomalies and pathological appearances such as consolidation or fibrosis. Transfer learning and self-configuring frameworks further accelerate model development by adapting pretrained weights and optimising hyperparameters for new datasets without extensive manual tuning. These methods support a range of clinical applications, including volumetric quantification of disease burden, surgical planning and radiomics studies, and have assumed particular importance in the context of emergent respiratory conditions such as COVID-19 and interstitial lung diseases.

Research from Nature Portfolio

Recent studies have developed a robust segmentation algorithm using a multi-resolution convolutional neural network trained on a polymorphic dataset comprising both human and animal CT scans. By incorporating specifically labelled and nonspecifically labelled examples, the model accurately segregates left and right lung regions across diverse pathologies including chronic obstructive pulmonary disease, confirmed COVID-19, lung cancer and idiopathic pulmonary fibrosis without disease-specific training data. Quantitative evaluation yielded an average Dice similarity coefficient of 0.985 and a symmetric surface distance below 0.5 mm. Subsequent lobar segmentation and hierarchical clustering of regional density patterns enabled the identification of four distinct COVID-19 radiographical phenotypes through analysis of lobar aeration and consolidation fractions.

Research from all publishers

A compact network paradigm has demonstrated that deep models with drastically reduced channel growth can achieve segmentation performance comparable to standard architectures while consuming up to 90 % less GPU memory and training significantly faster, thus facilitating deployment in resource-constrained settings. External validation of an nnU-Net-based model on contrast-enhanced chest CT scans from patients with pulmonary hypertension and interstitial lung disease achieved a mean Dice score of 0.990 and normalised surface distance of 0.983, with radiological review confirming minimal clinically relevant errors across diverse cohorts. Another approach combining a residual U-Net with a pre-trained ResNet-34 backbone on a cross-cohort dataset reported mean Dice coefficients above 0.93 for normal and pathological lungs, including cases with consolidation and pneumonia, underscoring the model’s generalisability across varying disease presentations and scanner types.

Deep Learning Methods in Lung CT Image Segmentation publication trend

The graph below shows the total number of articles in deep learning methods in lung ct image segmentation across all publications each year (not limited to Nature Index journals).

Technical terms

Convolutional neural network (CNN): A class of deep learning model that applies convolution operations to extract spatial features from images.

U-Net: A CNN architecture with a symmetrical encoder–decoder structure widely used for biomedical image segmentation.

nnU-Net: A self-adapting U-Net framework that automatically configures preprocessing, architecture and training for new segmentation tasks.

Dice similarity coefficient (DSC): A statistical measure quantifying the overlap between predicted and reference segmentations, expressed between 0 and 1.

Transfer learning: The technique of fine-tuning a pretrained model on a new dataset to leverage prior knowledge and reduce training time.

References

1. CT image segmentation for inflamed and fibrotic lungs using a multi-resolution convolutional neural network. Scientific Reports (2021).
2. Many Is Better Than One: An Integration of Multiple Simple Strategies for Accurate Lung Segmentation in CT Images. BioMed Research International (2016).
3. Pulmonary Lobe Segmentation With Probabilistic Segmentation of the Fissures and a Groupwise Fissure Prior. IEEE Transactions on Medical Imaging (2017).
4. Lung Segmentation on High-Resolution Computerized Tomography Images Using Deep Learning: A Preliminary Step for Radiomics Studies. Journal of Imaging (2020).
5. PocketNet: A Smaller Neural Network for Medical Image Analysis. IEEE Transactions on Medical Imaging (2023).
6. External validation, radiological evaluation, and development of deep learning automatic lung segmentation in contrast-enhanced chest CT. European Radiology (2023).
7. Lung Segmentation in CT Images: A Residual U-Net Approach on a Cross-Cohort Dataset. Applied Sciences (2022).