Automated quantification of lung pathology on micro-CT in diverse disease models using deep learning

Area of Science:

• * Biomedical imaging
• * Computational pathology
• * Preclinical research

Background:

• * Micro-computed tomography (micro-CT) is crucial for respiratory disease research, enhancing study efficiency and translatability to human trials.
• * Analyzing large micro-CT datasets is a significant bottleneck in preclinical respiratory disease research.
• * Current methods limit the efficient analysis of micro-CT data for longitudinal studies.

Purpose of the Study:

• * To develop a generic deep learning (DL) model for automated lung segmentation in rodent micro-CT images.
• * To assess the generalisability of the DL model across various lung pathologies, disease models, and scanner configurations.
• * To enable efficient and accurate quantification of lung pathology in preclinical respiratory disease studies.

Main Methods:

• * A deep learning (DL) based lung segmentation model was developed using longitudinal micro-CT images from diverse rodent studies.
• * 2D models were trained on axial, coronal, and sagittal slices, with predictions combined into a 2.5D model.
• * Model generalisability was tested on various studies (COVID-19, inflammation, fibrosis), rodent species, and scanner configurations, including public databases.

Main Results:

• * The 2.5D probability averaging DL model achieved a high mean Dice Similarity Coefficient (DSC) of 0.953 ± 0.023 on internal validation data.
• * The models demonstrated strong generalisability, with average DSC values from 0.89 to 0.94 across different conditions.
• * Biomarkers from automated segmentation closely agreed with manual segmentation, validating the model's ability to monitor disease progression and treatment response.

Conclusions:

• * The developed DL pipeline provides efficient analysis of large micro-CT datasets for lung pathology quantification.
• * The model is widely applicable across rodent disease models and various acquisition protocols.
• * This approach enables real-time insights into therapeutic efficacy in preclinical respiratory research.