ERS3062. Automated Deep Learning CT-Based Spleen Volume Segmentation: Defining Normal and Splenomegaly for Clinical Practice
Authors* Denotes Presenting Author
Alberto Perez *;
Mallinckrodt Institute of Radiology; University of Wisconsin School of Medicine and Public Health
Meghan Lubner;
University of Wisconsin School of Medicine and Public Health
John Garrett;
University of Wisconsin School of Medicine and Public Health
Ronald Summers;
National Institutes of Health
Perry Pickhardt;
University of Wisconsin School of Medicine and Public Health
Objective:
Imaging assessment for splenomegaly is not well defined and currently utilizes suboptimal unidimensional measures. Spleen volume provides a more direct measure for organ enlargement. We applied a validated deep learning artificial intelligence (AI) tool that automatically segments the spleen for organ volume, and sought to establish thresholds for splenomegaly.
Materials and Methods:
Spleen volumes were successfully derived with the deep learning tool from all 8853 asymptomatic outpatient adults (mean age 56.3 years; 4223M/4630F) who underwent CT for either colorectal screening (n=7688 unenhanced) or renal donor evaluation (n=1165, contrast-enhanced) and from 103 of 104 (>99%) adults with end-stage liver disease (ESLD) undergoing contrast enhanced CT for pre-transplantation evaluation. Linear regression analysis was utilized to assess major patient-specific determinate(s) of spleen volume amongst age, sex, height, weight, and BSA. Threshold for splenomegaly in the asymptomatic cohort was set at two standard deviations above the mean for the final modeled equation. Performance of the final modeled equation was assessed on the ESLD cohort. Accuracy of craniocaudal and maximal 3D linear measures was assessed. Unenhanced spleen volumes were standardized to a post-contrast equivalent, reflecting a small but constant 7.6% correction.
Results:
Mean standardized automated spleen volume was 216±100 ml in the asymptomatic cohort and demonstrated a normal distribution. Patient weight was the major determinant of spleen volume. From this, a linear weight-based splenomegaly threshold volume (in ml) = 3.0*(Wt [kg])+127. Above 125 kg, the volume cutoff was not increased. Linear measures demonstrated only moderate performance for identifying volume-defined splenomegaly within the asymptomatic cohort. For example, a craniocaudal threshold of 12 cm was 57% sensitive and 95% specific, and a maximal 3D linear threshold of 12 cm was 97% sensitive and 70% specific. Mean standardized automated spleen volume in the ESLD cohort was 796±457 ml with 85% (87 of 103) of patients meeting volume-defined splenomegaly.
Conclusion:
We derived a simple, weight-based threshold for splenomegaly using an automated spleen volume tool. If further validated in larger healthy and diseased cohorts, this approach could provide a more objective measure of spleen size from CT.
