Abstract 6315: Rapid automatic segmentation of murine lung volumes from micro-CT images using deep learning to reduce analysis burden

Micro-computed tomography (micro-CT) has emerged as a powerful tool for non-invasive small animal imaging, particularly in lung cancer and fibrosis research. This high-resolution imaging modality allows researchers to visualize and quantify lung structures and pathologies with unprecedented detail, enabling sophisticated longitudinal studies. However, the analysis of these complex 3D datasets presents a significant challenge. Manual segmentation of lung tissues from micro-CT images is time-consuming and prone to inter-observer variability, often becoming a bottleneck in study workflows. This analysis burden can impact study design, potentially limiting sample sizes or the frequency of imaging timepoints. As a result, there is a growing need for efficient, automated lung segmentation solutions that can accelerate analysis, improve reproducibility, and ultimately enhance the throughput and statistical power of preclinical lung studies. We address these challenges by training a new fully automatic deep learning model for whole lung segmentation. The goal of this work is to create a single model that generalizes well to data representing various manufacturers, disease states, and fields-of-view. In this pilot study, we collected 321 scans using a SkyScan 1276 (Bruker, Kontich, Belgium) and a Quantum GX2 (Revvity, Waltham, MA, USA). The lungs were segmented from each image by an expert reader. Data were shuffled and divided into groups for training, validation, and testing using a 70%-20%-10% split. Models were trained with different sets of hyperparameters and evaluated on the validation set. The model with the best validation Dice coefficient was selected for analyzing the test set. A linear regression analysis between the predicted and the ground truth test volumes resulted in a R2 value of 0.96, indicating strong correlation between the output of the model and the human annotator. The average time for the model to compute a single label is roughly 2 seconds, compared to an average of 15 minutes per volume for the expert reader, representing a significant improvement in processing time. Furthermore, the model is fully automatic, reducing the need for user input which might generate user fatigue. Taken together, these results suggest that this deep learning approach may be an effective aid for increasing throughput and reliability for micro-CT workflows in preclinical studies. Citation Format: Adam M. Aji, Thomas M. Kierski, Juan D. Rojas, Ryan C. Gessner, Yair Reisner, Irit Milman-Krentsis, John Hazle, Kiersten L. Maldonado, Charles Kingsley, Richard R. Bouchard, Tomasz J. Czernuszewicz. Rapid automatic segmentation of murine lung volumes from micro-CT images using deep learning to reduce analysis burden [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 6315.