Abstract 7444: Automated quantification of small animal liver tumor volume in longitudinal MR imaging using deep learning

Abstract In preclinical imaging studies with small animals, tracking tumor growth accurately over time can be challenging due to the time-consuming, labor-intensive, and error-prone nature of manual segmentation of 3D images. Orthotopic and metastatic cancer models can be particularly problematic as additional radiologic expertise is required to detect tumors within the context of deep anatomy, and individual animals may contain many nodules (ten or more) requiring many hours of segmentation effort. Therefore, the goal of this study was to develop and validate a deep learning model for automatic volumetric measurement of murine liver tumors in MR images to increase the practicality and adoptability of longitudinal imaging in preclinical research. We describe a method for automatically quantifying tumor volume over time using a dataset of 4.7T T2-weighted MR images taken from a study involving a genetic model of spontaneous liver cancer. The dataset included 894 images from 198 unique animals collected using a Biospec 47/40 (Bruker, Kontich, Belgium) over a span of 30 weeks. Each image was manually segmented by an expert annotator and tumor burden was tracked longitudinally to generate growth curves. A modified U-Net model built on the MONAI framework was trained to segment tumors. On a held-out test subset of 177 images (20% of the overall dataset), the model's predicted overall tumor burden showed a strong correlation with the human ground truth, achieving an R2 value of 0.88. The model processed each image in 0.34 seconds on a machine with an NVIDIA RTX 3090 GPU, significantly faster than the 5-10 minutes required for manual annotation. This efficiency translates to a substantial reduction in overall annotation time for studies with large datasets. This study demonstrates the feasibility of using deep learning for liver tumor segmentation in MR images. The proposed method offers a reliable and efficient alternative to manual segmentation, potentially improving the accuracy and speed of volumetric measurements in preclinical cancer research. Citation Format: Thomas M. Kierski, Adam M. Aji, Juan D. Rojas, Ryan C. Gessner, Laura Beretta, Suet Ying Kwan, John Hazle, Kiersten L. Maldonado, Charles Kingsley, Richard R. Bouchard, Tomasz J. Czernuszewicz. Automated quantification of small animal liver tumor volume in longitudinal MR imaging using deep learning [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 7444.