Clinical value of a CT radiomics model for predicting myocardial injury after cancer therapy

Technical Summary: Clinical Value of a CT Radiomics Model for Predicting Myocardial Injury After Cancer Therapy

Problem Statement
Early identification of patients at risk for cancer therapy-related myocardial injury remains a significant clinical challenge. Current surveillance relies heavily on echocardiography (e.g., Left Ventricular Ejection Fraction) and circulating biomarkers (e.g., troponins), which often detect injury only after myocardial damage has already occurred. While Cardiac Magnetic Resonance (CMR) offers higher tissue sensitivity, its cost and complexity limit its use as a universal pretreatment screening tool. Conversely, chest Computed Tomography (CT) is routinely acquired for cancer staging and surveillance but is underutilized for baseline cardiotoxicity risk stratification. This study addresses the gap in predicting myocardial injury before the initiation of cancer therapy using data already available in routine oncology workflows.

Methodology
This was a single-center, retrospective study conducted at Xuzhou Central Hospital involving patients with malignant tumors treated between January 2020 and December 2025.

• Cohort: 129 patients with complete imaging and outcome data were included in the radiomics workflow. For clinical inference and logistic regression, 128 patients were analyzed after excluding one case with missing covariates. The cohort consisted of 42 patients who developed myocardial injury and 86 controls.
• Outcome Definition: Myocardial injury was defined by the development of cardiovascular symptoms combined with at least one of: cardiac troponin elevation (>99th percentile), LVEF decrease ≥10% with LVEF <50%, or a relative reduction in Global Longitudinal Strain (GLS) >15%.
• Image Processing: Pretreatment chest CT images were resampled to an isotropic resolution of 1×1×1 mm³. Two experienced radiation oncologists manually delineated the left ventricular myocardial regions of interest (ROIs), excluding papillary muscles and blood pools.
• Feature Extraction & Selection: 128 radiomics features (first-order, shape, GLCM, GLRLM, GLSZM, GLDM, NGTDM) were extracted. Features were filtered through:
  1. Inter-observer reproducibility analysis (Intraclass Correlation Coefficient > 0.75), retaining 107 features.
  2. Variance threshold filtering (retaining 78 features).
  3. Spearman correlation analysis (|r| > 0.8) to remove redundancy (retaining 19 features).
  4. Elastic Net regularization for final selection, resulting in a 5-feature radiomics model.
• Model Development: Three models were constructed: a clinical model (based on Elastic Net-selected clinical variables), a radiomics model, and a combined model. The dataset was split 7:3 into training and test sets. Performance was evaluated using AUC, sensitivity, specificity, calibration (Hosmer-Lemeshow test), and Decision Curve Analysis (DCA).

Key Results

• Clinical Variables: In univariable analysis, alcohol use, hypertension, and absolute neutrophil count were associated with myocardial injury. However, in multivariable logistic regression, no clinical variable remained independently significant, though hypertension showed borderline significance (P=0.059).
• Radiomics Features: The final radiomics model retained five features: 10Percentile (first-order), Flatness, LeastAxisLength, Maximum2DDiameterColumn (shape), and SizeZoneNonUniformity (texture).
• Predictive Performance:
  • Radiomics Model: Achieved an AUC of 0.867 (95% CI: 0.722–0.972) in the test set, with a sensitivity of 0.692 and specificity of 0.962 at the optimal cutoff.
  • Clinical Model: Achieved an AUC of 0.601.
  • Combined Model: Achieved an AUC of 0.805.
  • Comparison: The radiomics model demonstrated significantly better discrimination than the clinical model (ΔAUC = 0.266, P = 0.036). The difference between the radiomics and combined models was not statistically significant (P = 0.475).
• Calibration & Utility: The radiomics model showed acceptable calibration (Hosmer-Lemeshow P = 0.066). Decision Curve Analysis indicated a higher net benefit for the radiomics model compared to "treat-all" or "treat-none" strategies across a threshold probability range of 0.01–0.60.

Key Contributions

1. Pretreatment Risk Stratification: The study demonstrates that quantitative myocardial phenotypes extracted from routine pretreatment chest CT can identify patients at increased risk for subsequent therapy-related myocardial injury, shifting risk assessment earlier in the care pathway.
2. Superiority of Imaging over Clinical Factors: The radiomics model significantly outperformed a model based solely on clinical variables, suggesting that pretreatment myocardial heterogeneity (captured via gray-level distribution, geometry, and texture) contains predictive information not reflected in standard clinical history or biomarkers.
3. Workflow Integration: By utilizing existing CT data acquired for oncology staging, the approach offers a low-burden, non-invasive method for baseline cardiotoxicity assessment without requiring additional scans or significant workflow disruption.

Significance and Claims
The authors claim that this study provides preliminary evidence that a pretreatment chest CT-based myocardial radiomics model has potential for identifying high-risk patients. The paper emphasizes that while the radiomics model showed strong discrimination, the findings are exploratory. The authors explicitly state that the approach supports "low-burden baseline cardiotoxicity risk assessment" using data already acquired during routine care.

However, the paper maintains a modest tone regarding clinical implementation. It acknowledges that the study is retrospective, single-center, and relatively small, lacking an independent external validation cohort. Consequently, the authors conclude that larger external cohorts are required for validation before the model can be implemented clinically. They note that the combined model's lack of significant improvement over the radiomics model in the test set does not rule out the value of clinical data, but rather suggests that larger studies are needed to confirm potential complementarity. The study positions itself as a step toward integrating quantitative imaging phenotypes into the broader understanding of therapy-related myocardial vulnerability.