Deep Learning Segmentation of Triple-Negative Breast Cancer (TNBC) Patient Derived Tumor Xenograft (PDX) and Sensitivity of Radiomic Pipeline to Tumor Probability Boundary

Simple Summary Co-clinical trials are an emerging area of investigation in which a clinical trial is coupled with a corresponding preclinical trial to inform the corresponding clinical trial. The preclinical arm aids in assessing therapeutic efficacy, patient stratification, and designing optimal imaging strategies. There is much interest in harmonizing preclinical and clinical quantitative imaging pipelines. Radiomics is widely explored in clinical imaging to predict response to therapy. In preclinical imaging, high-throughput radiomic analysis is limited by manual delineation of tumor boundaries, which is labor intensive with poor reproducibility. Our proposed deep-learning-based system was trained to automatically segment tumors from multi-contrast MR images and extract radiomic features. The proposed method is highly reproducible with significant correlation in radiomic features. The deployment of this pipeline in the preclinical arm would provide high throughput and reproducible radiomic analysis. Preclinical magnetic resonance imaging (MRI) is a critical component in a co-clinical research pipeline. Importantly, segmentation of tumors in MRI is a necessary step in tumor phenotyping and assessment of response to therapy. However, manual segmentation is time-intensive and suffers from inter- and intra- observer variability and lack of reproducibility. This study aimed to develop an automated pipeline for accurate localization and delineation of TNBC PDX tumors from preclinical T1w and T2w MR images using a deep learning (DL) algorithm and to assess the sensitivity of radiomic features to tumor boundaries. We tested five network architectures including U-Net, dense U-Net, Res-Net, recurrent residual UNet (R2UNet), and dense R2U-Net (D-R2UNet), which were compared against manual delineation by experts. To mitigate bias among multiple experts, the simultaneous truth and performance level estimation (STAPLE) algorithm was applied to create consensus maps. Performance metrics (F1-Score, recall, precision, and AUC) were used to assess the performance of the networks. Multi-contrast D-R2UNet performed best with F1-score = 0.948; however, all networks scored within 1-3% of each other. Radiomic features extracted from D-R2UNet were highly corelated to STAPLE-derived features with 67.13% of T1w and 53.15% of T2w exhibiting correlation rho >= 0.9 (p <= 0.05). D-R2UNet-extracted features exhibited better reproducibility relative to STAPLE with 86.71% of T1w and 69.93% of T2w features found to be highly reproducible (CCC >= 0.9, p <= 0.05). Finally, 39.16% T1w and 13.9% T2w features were identified as insensitive to tumor boundary perturbations (Spearman correlation (-0.4 <= rho <= 0.4). We developed a highly reproducible DL algorithm to circumvent manual segmentation of T1w and T2w MR images and identified sensitivity of radiomic features to tumor boundaries.

Automated summary

Co-clinical trials combine clinical and preclinical trials to support corresponding clinical trials, with radiomics being widely explored in clinical imaging for predicting therapy response. This study developed a reproducible DL algorithm for automated segmentation and feature extraction from preclinical MR images. The D-R2UNet network architecture demonstrated high correlation and reproducibility with STAPLE-derived features, providing a sensitive analysis of radiomic features to tumor boundaries.