Abstract 7443: Feasibility of automatic spleen volumetry from 3D ultrasound images of mice using deep learning

Abstract The spleen plays a critical role in immune function and blood filtration. Interestingly, the spleen's volume changes in response to various diseases. For example, enlargement of the spleen (splenomegaly) can arise because of cancer (solid and liquid tumors), liver diseases (MASH), Lupus, or various infections. Conversely, atrophy or shrinking of the spleen can arise from conditions such as graft-versus-host-disease, inflammatory bowel, or sickle cell disease. Though the size of the spleen is a relatively well-established biomarker that may be useful in several preclinical studies, it is currently difficult and costly to measure for large cohorts of mice over multiple timepoints. It is possible to measure spleen volume and weight ex vivo at a terminal timepoint, but this approach is not amenable to longitudinal tracking. Alternatively, imaging using ultrasound or magnetic resonance can be used to interrogate the spleen noninvasively and repeatedly, but this method requires data to be segmented in 3D. Segmentation often creates a bottleneck in study design because it is time consuming. Furthermore, measurements derived from human-annotated data are prone to inter-user variability and user error. Towards addressing these challenges with measuring spleen volume in longitudinal preclinical studies, we trained an AI model to segment spleens from a diverse set of 3D ultrasound images. In total, 180 images were acquired in mice from a Vega ultrasound scanner (Revvity, Waltham, MA, USA). These images were varied in terms of animal strain, positioning, field of view, and disease state with splenic volumes ranging between approximately 20 mm3 and 150 mm3. These data included samples from healthy controls and mice with conditions including solid tumors, blood cancer, Lupus, and graft versus host disease which were annotated by expert human readers. Data were shuffled and split into training and validation groups with an 80-20 split and the training set was used to train a modified U-Net model. Inference on the validation set after training yielded a Dice coefficient of 0.86, and the total processing time for the 36 images was 49.0 seconds (1.36 sec/image) on an NVIDIA RTX 3090 GPU. A linear regression analysis between the ground truth volumes and the volumes predicted by the model resulted in a R2 value of 0.85. These results suggest that this problem is an excellent candidate for automation using artificial intelligence, particularly with an automated ultrasound scanner such as the Vega where positioning is highly reproducible. The training set size (N=144 images) is quite small by conventional deep learning standards, and we expect the segmentation accuracy to increase further as more images are added to the training pool. Future work will focus on collecting more in vivo images from a variety of murine models. Citation Format: Adam M. Aji, Thomas M. Kierski, Juan D. Rojas, Bincy John, David Harris, Ryan C. Gessner, Craig McMannus, Jeffrey Peterson, Tomasz J. Czernuszewicz. Feasibility of automatic spleen volumetry from 3D ultrasound images of mice 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 7443.