Co-clinical trials are an emerging research design in which a clinical cancer trial runs in parallel with a matched preclinical trial using patient-derived xenografts (PDXs) -- tumors taken from human patients and grown in mice. PDX models better capture the heterogeneity of real human cancers than traditional cell lines, making them more predictive of treatment outcomes in patients. The preclinical arm can rapidly test biomarkers, treatment combinations, and imaging strategies that would take years to evaluate in humans.
MRI is a critical tool in these co-clinical pipelines, providing detailed images of tumor size, shape, and texture across treatment timepoints. However, every quantitative analysis step first requires tumor segmentation -- precisely outlining the tumor boundary in each image slice. Done manually, segmentation is time-consuming, labor-intensive, and produces different results depending on which expert does it, or even when the same expert repeats the task. This variability limits reproducibility and hinders high-throughput analysis.
Beyond accurate tumor delineation, radiomics -- the extraction of hundreds of quantitative features (shape, intensity, texture) from segmented tumor regions -- offers a path to characterizing tumor heterogeneity and predicting treatment response without additional tests. However, radiomic features vary depending on how precisely the tumor boundary is drawn. Understanding which features are robust to boundary uncertainty is essential before using them as biomarkers.
This study addresses both problems simultaneously: automating segmentation with deep learning to replace manual delineation, and systematically identifying which radiomic features remain stable even when tumor boundaries are imperfectly defined.
The study used TNBC PDX tumors implanted in the inguinal mammary fat pad of 49 mice. Each mouse was imaged using a 7-Tesla small animal MR/PET scanner at Washington University, acquiring both T1-weighted (T1w) and T2-weighted (T2w) MRI sequences. T1w images highlight tissue anatomy with strong signal from fat and contrast-enhancing regions; T2w images are sensitive to fluid and soft tissue and are typically more informative for tumor boundary delineation.
Four expert radiologists independently segmented each tumor, and the STAPLE algorithm (Simultaneous Truth and Performance Level Estimation) was applied to their four sets of annotations to create a single consensus segmentation map. STAPLE uses expectation-maximization to estimate the probability that each voxel is part of the tumor, accounting for the variability between annotators rather than simply averaging. This consensus map served as the ground truth against which automated methods were compared.
Five deep learning architectures were implemented and compared: standard U-Net, Dense U-Net, Residual U-Net (Res-Net), Recurrent Residual U-Net (R2UNet), and Dense Recurrent Residual U-Net (D-R2UNet). All are variants of the encoder-decoder architecture where the network progressively compresses the image into abstract features (encoding), then reconstructs a pixel-level segmentation map (decoding). The key differentiators among variants are: residual connections that mitigate the vanishing gradient problem in deep networks; recurrent convolutional layers that refine feature representations through multiple passes; and dense connections that allow every layer to directly access features from all previous layers, promoting feature reuse.
Training used 41 mice (255 MR slices) with fivefold cross-validation, and testing used 8 held-out mice (39 slices). Data augmentation (rotation, shifting, shearing) mitigated overfitting given the small dataset. Multi-contrast input combined T1w and T2w as two channels of the same tensor. Models were trained for 250 epochs with the Adam optimizer using dice loss as the training objective.
All five networks performed better than previously published preclinical tumor segmentation approaches, including a prior k-means level-set method (F1-score 0.82) and a sarcoma U-Net (F1-score 0.84). Among the five architectures, D-R2UNet with multi-contrast input achieved the best performance: F1-score of 0.948, precision of 0.970, recall of 0.928, and AUC of 0.963. Notably, all five networks scored within 1-3% of each other, suggesting the overall U-Net family is robust for this task regardless of which variant is used.
Multi-contrast input (T1w + T2w combined) consistently outperformed T2w-only input. The best T2w-only model (Dense R2U-Net) achieved an F1-score of 0.935, compared to 0.948 for the same architecture with both MRI contrasts. This mirrors findings from brain segmentation studies where multi-contrast data allows the network to leverage complementary tissue properties from different pulse sequences.
Tumor volume agreement between the D-R2UNet and the STAPLE consensus was very high: concordance correlation coefficient (CCC) of 0.991, and Spearman correlation of 0.99 (p less than 0.001). Bland-Altman analysis showed a mean volume bias of only 4.6%, meaning the automated algorithm slightly underestimated tumor volume relative to STAPLE. The bias decreased as tumor volume increased -- larger tumors were measured more accurately, likely because boundary uncertainty matters less as a fraction of the total volume.
For reproducibility (test-retest reliability), D-R2UNet outperformed every individual expert and the STAPLE ensemble, achieving a mean volume bias of just 1.02% between two training runs. Experts, by contrast, showed biases ranging from -2.2% to 8.37% between repeated measurements of the same scans, and even the STAPLE consensus showed a 2.8% bias. This demonstrates that the algorithm is more consistent than human annotators.
After segmentation, 144 radiomic features were extracted from each T1w and T2w volume, spanning morphological (shape), statistical, histogram, and multiple texture matrix categories (GLCM, GLRLM, GLSZM, GLDZM, NGLDM, NGTDM). The key question was: how well do features extracted from automated D-R2UNet segmentations agree with features extracted from the STAPLE expert consensus?
Overall, 67.13% of T1w features and 53.15% of T2w features showed a Spearman correlation of 0.9 or higher between D-R2UNet and STAPLE maps. All 12 morphological (shape-based) features showed strong correlation (0.83 to 1.0). Global intensity features showed near-perfect correlation for T1w (0.95-1.0) and strong-to-moderate correlation for T2w (0.66-1.0). Texture features showed the widest variation, because they depend on intensity quantization -- and changing the tumor boundary by even a few voxels can alter the intensity distribution and therefore the quantization levels used to compute texture metrics.
For reproducibility, 86.71% of T1w features and 69.93% of T2w features extracted from D-R2UNet had a concordance correlation coefficient (CCC) of 0.9 or higher. This outperformed the expert STAPLE reproducibility on several metrics, consistent with the lower volume bias observed for the algorithm. Greater than 80% of morphological features and greater than 90% of statistical features were highly reproducible for both D-R2UNet and STAPLE methods.
The consistently higher reproducibility of T1w features compared to T2w features is explained by texture heterogeneity: T2w images have more inherent texture variation, so even small boundary changes have disproportionate effects on texture-based radiomic metrics that rely on intensity binning within the segmented region.
A key analysis asked: which radiomic features are insensitive to small perturbations in the tumor boundary? This matters because any segmentation method -- automated or manual -- has some uncertainty at the edges. Features that change dramatically when the boundary shifts by a few voxels are poor biomarker candidates; features that remain stable despite boundary variation are more trustworthy.
Sensitivity was measured as the Spearman correlation between the change in feature value and the change in tumor volume between STAPLE and D-R2UNet maps. Features with correlation between -0.4 and +0.4 were classified as boundary-insensitive. The result: 95 T1w features (39.16%) and 50 T2w features (13.9%) were robust to boundary perturbations. Hierarchical clustering organized these features into 23 (T1w) and 26 (T2w) independent clusters, helping identify which groups of related features can be represented by a single robust representative.
Of these boundary-robust features, a further subset was identified as meeting all three quality criteria simultaneously: high correlation with STAPLE (Spearman rho at least 0.9), high reproducibility (CCC at least 0.9), and insensitivity to boundary perturbation. This triple-filtered subset comprised 56 T1w features (36.16%) and 20 T2w features (13.9%). These are the features recommended for use as imaging biomarkers in co-clinical studies.
This analysis directly addresses a major challenge in radiomic research: the lack of standardization around which features are reliable enough to serve as biomarkers. By providing a data-driven, quantitative filter, the study creates an evidence base for selecting trustworthy radiomic biomarkers in the TNBC PDX preclinical context.
The most important contribution of this work is enabling high-throughput, reproducible radiomic analysis in the preclinical arm of co-clinical trials. Previously, the bottleneck was the manual tumor delineation step: each mouse required multiple expert hours to annotate, limiting how many animals could be analyzed and introducing variability that reduced confidence in downstream results. The automated pipeline removes this bottleneck entirely.
The D-R2UNet's dense residual and recurrent convolutional architecture proved particularly effective for the small dataset (49 mice, 255 training slices) by maximizing feature reuse through dense connections and learning iteratively refined representations through recurrent layers. These architectural choices help the network learn from fewer examples without overfitting.
The finding that texture-based features are more sensitive to boundary uncertainty than shape or intensity features has practical implications: researchers using radiomic biomarkers from preclinical MRI should prioritize shape and intensity features over texture features unless their segmentation method has been validated for boundary precision. The hierarchical clustering additionally provides a tool for reducing feature redundancy when selecting a biomarker panel.
The pipeline is publicly available on GitHub and the Washington University Co-Clinical Imaging Research Resource web portal, making it directly usable by other research groups running TNBC PDX co-clinical trials. This open-source release multiplies the impact by enabling other labs to immediately benefit from the validated pipeline without rebuilding it from scratch.
The D-R2UNet deep learning pipeline accurately segments TNBC PDX tumors from multi-contrast preclinical MRI with an F1-score of 0.948, substantially outperforming prior automated methods and matching expert consensus with greater test-retest reproducibility than human annotators. This makes it a viable replacement for manual segmentation in the preclinical co-clinical trial pipeline.
The radiomic analysis identified a curated subset of boundary-robust, highly reproducible features that correlate well with expert-derived measurements. These features provide a validated starting point for preclinical radiomic biomarker discovery -- a necessary step toward harmonizing radiomic analysis between preclinical and clinical imaging in co-clinical trial design.
The long-term clinical relevance of this work lies in translational harmonization: co-clinical trials can only inform clinical trial design if the imaging biomarkers measured in mouse models and in humans are comparable. Automated, reproducible segmentation pipelines applied consistently across both arms of a co-clinical trial are a prerequisite for this alignment.
Future work would extend this pipeline to contrast-enhanced MRI and PET imaging data, increase the training dataset size, and prospectively validate the identified biomarker features as predictors of response to neoadjuvant chemotherapy in both the preclinical and clinical arms of co-clinical TNBC trials.