Chemotherapy Response Score (CRS) Overview

• Chemotherapy Response Score (CRS) is a histopathological metric that stratifies post-NACT tumor specimens into three tiers based on residual disease and fibro-inflammatory response.
• It underpins advanced radiomics pipelines integrating CT imaging, semi-automated lesion segmentation, and robust feature selection to predict chemotherapy outcomes.
• The integration of machine learning models with meticulous robustness assessments offers actionable insights for personalized treatment in high-grade serous ovarian carcinoma.

The Chemotherapy Response Score (CRS) is a three-tier histopathological metric designed to evaluate tumor response in omental specimens following neoadjuvant chemotherapy (NACT) for high-grade serous ovarian carcinoma (HGSOC). The score stratifies tumors based on the degree of residual disease and fibro-inflammatory response, and has been operationalized as a key clinical endpoint in multiple radiomics and predictive modeling studies employing pre- and post-NACT computed tomography (CT) data. Recent research couples CRS with advanced radiomics and automated feature robustness assessment, enabling the development of reliable machine learning models for non-invasive chemotherapy response prediction (Hatamikia et al., 13 Jan 2026).

1. Definition and Clinical Basis of Chemotherapy Response Score

The CRS metric, as specified by Böhm et al., partitions post-NACT omental tumor specimens into distinct response strata, emphasizing histopathological characterization:

CRS Tier	Histopathologic Criteria
CRS 1	Abundant residual tumor with no or minimal response
CRS 2	Appreciable fibro-inflammatory response but significant viable tumor remains
CRS 3	Complete or near-complete response: no tumor or only minimal (<2 mm) residual tumor nests

In predictive model development, CRS 1–2 are commonly grouped as "non-complete response," while CRS 3 represents "complete response." This dichotomization provides a clinically actionable endpoint for assessing efficacy of NACT and for guiding post-surgical therapy selection (Hatamikia et al., 13 Jan 2026).

2. Radiomics Pipeline for CRS Prediction

Radiomics analysis supporting CRS prediction is predicated on a standardized pipeline comprising image acquisition, lesion segmentation, feature extraction, and robust feature selection:

• Image Acquisition: Pre- and post-NACT contrast-enhanced abdominopelvic CT scans from institutions such as Cambridge and Barts, adhering to standard clinical protocols.
• Image Pre-processing:
  • Lesion segmentation is executed semi-automatically (Microsoft InnerEye), with radiologist oversight.
  • Intensity discretization employs a fixed bin width of 4 Hounsfield Units; no additional intensity normalization or resampling is performed.
• Region of Interest (ROI) Specification:
  • Lesions are labeled by anatomical site (omental, pelvic).
  • Two pooling strategies: "largest lesion only" and "merged" (all lesions combined).
  • Peripheral rim ROIs (6 mm) are generated for rim radiomics in CRS models.
• Feature Extraction:
  • 102 features per ROI derived with Pyradiomics v3.0.1, spanning shape descriptors (e.g., sphericity, surface area), first-order statistics (mean, entropy, skewness, kurtosis), and multiple texture matrices (GLCM, GLRLM, GLSZM, GLDM, NGTDM).

This pipeline underpins the extraction of quantifiable markers potentially insensitive to reader or institution-specific protocol differences (Hatamikia et al., 13 Jan 2026).

3. Robustness Assessment and Feature Selection

A distinguishing aspect of recent CRS prediction work is the explicit modeling of inter-observer segmentation variability via automated random perturbation of VOIs (volumes of interest):

• Robustness Modeling: Randomized segmentation, applying small geometric perturbations, mimics plausible variability between human annotators.
• Intraclass Correlation Coefficient (ICC) Computation: For each radiomics feature $f$, the ICC is computed between original and perturbed VOIs using the two-way consistency model (single measurement):

$$\mathrm{ICC}(f) = \frac{\sigma^2_{\text{between}}}{\sigma^2_{\text{between}} + \sigma^2_{\text{within}}}$$

• ICC Interpretation:
  • ICC ≥ 0.90: "excellent"
  • ICC 0.70–0.90: "moderate"
  • ICC < 0.70: "poor"

Feature selection methodologies include:

• Filtering: Univariate Feature Selection (UFS) within 5-fold stratified CV removes highly correlated/non-informative features.
• Feature Selection Algorithms: F-score, Relief, Mutual Information, Gini importance, LASSO, Genetic Algorithm (GA), Sequential Forward/Backward Search (SFS/SBS), Recursive Feature Elimination (RFE).
• PREDICTIVE&ROBUST Approaches:

1. Fully robust: Pre-filter features with ICC > 0.80.
2. Semi-robust: At each iteration, ≥ 80% of features must have ICC > 0.80.
3. Weighted robustness: Composite relevance-robustness score $s' = (1-w) s + w c$, with $w=0.5$ (where $s$ is predictive relevance, $c$ is mean ICC).

These steps maximize model stability and reliability by deprioritizing features sensitive to segmentation noise (Hatamikia et al., 13 Jan 2026).

4. Machine Learning Models for CRS Prediction

Modeling to predict CRS employs classical linear classifiers, with feature selection strategies that incorporate robustness constraints explicitly:

• Classifiers:
  • Logistic Regression (LR)
  • Linear Discriminant Analysis (LDA)
• Training Strategy:
  • 5-fold stratified cross-validation on the training set (OV04) to optimize FS + classifier combinations, using AUC as the selection criterion.
  • No additional hyperparameter grid search beyond intrinsic LASSO/GA settings.
• Robustness-Constrained Feature Sets: Only feature sets selected per the PREDICTIVE&ROBUST criteria are used for final model training.

The integration of linear models with rigorously filtered and robustness-weighted feature sets aims to provide both interpretability and reproducibility, essential for biomarker development in translational imaging (Hatamikia et al., 13 Jan 2026).

5. Performance Evaluation and Lesion-Specific Insights

The benchmark for CRS model evaluation is external validation using an independent cohort (BARTS):

Lesion Site & ROI Strategy	CRS AUC	G-Mean	Sensitivity	Specificity
Omental (merged, weighted, SFS+LDA)	0.77	0.72	0.56	0.92
Omental (rim radiomics only)	0.74	—	—	—
Pelvic (rim radiomics only)	0.42	—	—	—

• Omental Lesion Dominance: Omental lesions consistently provided the highest AUC (0.77 vs. 0.62–0.68 for pelvic/all), with a favorable balance of sensitivity and specificity.
• Robustness of Omental Features: A greater proportion of omental features exhibited "excellent" ICC, indicating superior measurement stability under segmentation perturbation.
• Rim Radiomics: For omental disease, rim-based features matched whole-ROI performance (AUC = 0.74), but rim-only pelvics showed marked performance degradation (AUC = 0.42).
• Interpretation: These results suggest that omental tumor imaging yields more reproducible texture and shape signals that correlate strongly with histopathologic treatment response, and that model robustness is lesion site-dependent.

No confidence intervals were reported for these metrics (Hatamikia et al., 13 Jan 2026).

6. Clinical Implications, Limitations, and Prospects

Incorporating segmentation robustness in feature selection demonstrably improves the reliability of radiomics-derived CRS predictors, thereby supporting clinical deployment for stratifying patients based on predicted NACT response:

• Clinical Validity: Robust radiomics models may triage patients unlikely to benefit from NACT, guiding more effective personalized therapy.
• Methodological Constraints:
  • Semi-manual segmentation is labor-intensive and limits scalability.
  • CT acquisition variability (e.g., slice thickness, scanner heterogeneity) was not comprehensively harmonized.
  • No prospective, real-time deployment or validation has been performed.
• Future Directions:
  • Automating lesion delineation via deep learning to enable on-the-fly generation of perturbed VOIs for robustness modeling.
  • Integration of real-time radiomics workflows in multidisciplinary cancer care settings.
  • Extension to multi-parametric or "habitat" approaches to enhance model generalizability across institutions.

A plausible implication is that the described framework—by simulating inter-observer segmentation variability and prioritizing robust features—lays the foundation for reproducible CT-based biomarkers, facilitating objective and individualized patient management in ovarian cancer. The strategy outlined offers a blueprint for future studies targeting robust predictive radiomics in other cancer types and contexts (Hatamikia et al., 13 Jan 2026).