Deep Learning Based Tumor Segmentation on MRI of Prostate Cancer Patient-Derived Xenografts in Mouse Models

Motivation: Tumor volume measurement in non-subcutaneous murine xenografts is labor-intensive and prone to variability, necessitating automated approaches to improve efficiency and accuracy. Goal(s): Develop an automated deep-learning pipeline for accurately measuring xenograft tumor volumes from MRI scans, minimizing human intervention and improving reproducibility. Approach: We designed a two-step pipeline involving ResNet-50 for tumor classification and D-R2UNet for segmentation. Models were trained on MRI datasets with 5-fold cross-validation. Results: The classifier achieved 89.9% accuracy. The D-R2UNet demonstrated superior segmentation performance across anatomical sites, significantly reducing segmentation time while maintaining high accuracy. Impact: This automated segmentation pipeline enhances efficiency in preclinical tumor studies, reducing manual effort and interuser variability. It provides a robust tool for evaluating treatment efficacy, potentially enabling broader use in diverse xenograft studies and informing translational research.