Dynamic Segmentation of Pre-Clinical Imaging Using Deep Learning
Abstract
Purpose
Three-dimensional multi-organ segmentation is a critical step in preclinical micro-CT imaging for quantitative analysis and reproducibility of longitudinal studies. However, manual delineation remains time-consuming and prone to inter-operator variability, while existing automatic methods often lack robustness when facing anatomical variability and heterogeneous acquisition conditions. The purpose of this work is to investigate whether a lightweight 3D U-Net, combined with a carefully designed learning pipeline, can achieve competitive multi-organ segmentation performance without relying on complex architectures.
Methods
A public whole-body mouse micro-CT dataset including longitudinal contrast-enhanced acquisitions and expert multi-organ manual segmentations was used. Volumes were resampled to an isotropic spatial resolution, intensity-normalized using Z-score normalization, and divided into overlapping 3D patches of size 128×128×128 voxels. A lightweight 3D U-Net was trained using a composite loss function combining Dice loss and weighted cross-entropy. Several data augmentation strategies (noise, brightness, and contrast variations, applied individually or in combination) as well as different pipeline configurations were systematically evaluated. Segmentation performance was assessed organ-wise using the Dice similarity coefficient on independent validation and test sets.
Results
The proposed approach achieved stable and high segmentation performance across most evaluated organs. The results demonstrate that preprocessing quality and data augmentation strategy selection have a greater impact on segmentation performance than model architectural complexity. Certain augmentation combinations significantly improved inter-subject robustness, whereas overly aggressive augmentation degraded performance for specific anatomical structures.
Conclusion
A lightweight 3D U-Net integrated within a carefully optimized learning pipeline enables reliable and reproducible multi-organ segmentation in preclinical micro-CT imaging. This approach provides an efficient and practical solution well suited to preclinical imaging workflows and transferable to other medical imaging applications.