From 2D to 3D: Automated ultrasound segmentation and cross-sectional validation in murine tumor models

Ultrasound (US) is a widely used method for non-invasive tumor monitoring. Semantic segmentation of tumors in US imagery is a necessary step to reconstruct 3D geometry of a region of interest (ROI). Still, the segmentation task remains challenging due to variable signal-to-noise ratio (SNR) and modest soft-tissue contrast in US imaging. Our objective was to generate a new dataset of murine tumors and present a standardized pipeline for 3D volume reconstruction, suitable for preclinical research.

Human LN229 and murine 4T1, PanO2, B16 and LLC tumors were imaged in vivo using high-frequency US systems (Vevo F2, Vevo 2100). Our dataset comprised 3442 images, including expert-curated masks and a challenging out-of-distribution (OOD) test set. We evaluated U-Net, Res U-Net, Attention U-Net, and R2AU-Net, with and without autoencoder pretraining on unlabeled frames. For volumetry, 2D masks were converted to voxel grids using known imaging geometry; through-plane interpolation and Marching Cubes surface extraction enabled shape-agnostic 3D volume computation. We compared US-derived volumes with micro-CT, calipers, mold-based, and tumor weight.

Across random train-test splits, all models achieved Dice > 0.90. On the subject-independent special testing dataset, performance decreased, indicating limited generalization under distribution shift; the pretrained Attention Res U-Net achieved the highest overlap (Dice 0.750, IoU 0.604), while the pretrained Attention U-Net also remained comparatively robust (Dice 0.731). The 3D reconstruction pipeline produced consistent longitudinal volumes, and cross-modal comparison showed that US-based volumes agreed with micro-CT and physical measurements.

This study presents a standardized workflow for automated tumor segmentation and 3D ultrasound-based volumetry, enabling more objective and reproducible assessment of tumor burden in preclinical oncology studies.