Evaluation and Optimization of AI-Based Organ Segmentation Algorithms with Whole-Body CT Data
Abstract
Purpose
AAPM Task Group 430 aims to comprehensively evaluate organ and effective dose (ED) estimations from whole body CT data to support recommendations for DLP-to-ED conversions factors and their associated uncertainties. Accurate whole-body organ segmentation is critical for this task. However, commercial tools exhibit significant inter-software variability, and open-source models (e.g., TotalSegmentor, Vista3D) fail on whole-body (head-to-toe) datasets due to limitation in their initial training. This study introduces a "Split-Merge" workflow to optimize open-source AI models, compares performance with a new model from NeuralRad, and establishes a framework for open data dissemination.
Methods
A cohort of 120 whole-body patient CT scans was used to develop a robust preprocessing pipeline combining iterative 2D/3D traditional segmentation and multi-parameter connectivity analysis to remove extraneous non-anatomical structures (tables, wires). Field-of-view limitations of existing AI models were addressed using a "Split-Merge" algorithm that partitions the whole-body volume into overlapping Head, Upper-Body, and Lower-Body sections, executes deep learning model predictions on each section, and recombines sections via union logic, implemented within a GUI-based platform. Performance was evaluated via visual inspection and quantitative comparisons of the Voxel Assignment Percentage (VAP).
Results
The "Split-Merge" workflow significantly reduced truncation artifacts compared to open-source AI models. Visual comparison of 2D slices and 3D renderings confirmed that while optimized open-source models capture major organs, they frequently miss muscle and subcutaneous tissues. NeuralRad achieved ~100% VAP (indicating complete soft-tissue capture), whereas the VAP for optimized open-source models remained low.
Conclusion
A robust pipeline for whole-body organ segmentation was developed and shown to mitigate the limitations of commercial and generic open-source AI solutions. While "Split-Merge" enables the use of open-source models for whole-body segmentation, NeuralRad demonstrated superior anatomical completeness required for accurate effective dose estimation in CT. Future work will focus on ground-truth validation and dissemination of this dataset to the community.