TL;DR
This paper introduces a novel "radiomic signature" (RadSig) derived from FDG-PET scans that can predict how well triple-negative breast cancer (TNBC) patients will respond to chemotherapy. By optimizing these radiomic features in preclinical models (patient-derived xenografts or PDXs) before applying them to human patients, they achieved superior prediction accuracy compared to traditional SUV metrics, paving the way for more personalized cancer treatment.
What Problem Does This Paper Address?
Triple-negative breast cancer (TNBC) is an aggressive form of breast cancer with poor outcomes. Predicting response to neoadjuvant chemotherapy (NAC) is crucial to avoid ineffective treatments and tailor therapy. However, TNBC is highly heterogeneous, making it difficult to predict treatment response. Traditional imaging metrics like SUV (Standardized Uptake Value) have limitations. Radiomics, which extracts quantitative features from medical images, offers a promising avenue, but its application in a co-clinical setting—where preclinical models closely mimic human tumors—to optimize predictive models for clinical use is still an emerging area. This paper aims to bridge this gap by developing and validating a robust radiomic signature for predicting NAC response in TNBC using a co-clinical FDG-PET imaging approach.
Methods
The study employed a co-clinical design, linking patient data with matched preclinical models (PDXs).
- Preclinical Optimization (PDX):
- TNBC PDX models were generated from patient tumors.
- FDG-PET/CT imaging was performed on these PDXs.
- Radiomic Feature Extraction: 131 imaging features were extracted from segmented tumors, adhering to IBSI guidelines.
- Robustness Assessment: Features were evaluated for:
- Reproducibility (Test-Retest): Assessed using Lin's Concordance Correlation Coefficient (LCC) on repeated scans of the same PDX.
- Cross-correlation: Spearman correlation was used to identify highly correlated features, which were then reduced.
- Volume Dependency: Features were regressed against tumor volume to identify and potentially mitigate volume-induced bias.
- Feature Selection & Model Training: A subset of robust features was selected using the ReliefF algorithm. Machine learning (ML) algorithms (CART, Naïve Bayes, SVM) were trained on these features to predict/assess response to therapy in PDXs. The top four features for prediction (baseline) and assessment (change from baseline) formed the "radiomic signature" (RadSig).
- Clinical Validation:
- The PDX-optimized RadSig was then applied to data from a clinical trial involving TNBC patients undergoing NAC.
- FDG-PET/MR imaging was performed at baseline and during treatment.
- The performance of RadSig in predicting and assessing response in the clinical arm was compared to traditional SUV metrics (SUVmean, SUVmax, SULpeak).
Datasets
- Preclinical Dataset: Data from TNBC patient-derived tumor xenografts (PDX) in mice. The exact number of PDX models used for feature optimization is not explicitly stated but is implied to be substantial for robust feature selection.
- Clinical Dataset: Data from 20 newly diagnosed Stage II or III TNBC patients recruited into an ongoing co-clinical trial. This dataset was used for an interim analysis to validate the PDX-optimized RadSig.
Key Insights and Ideas
- Co-clinical approach is powerful: Using PDXs as a preclinical platform allowed for the optimization of radiomic features that are robust and predictive, which were then successfully translated to the clinical setting. This highlights the value of bridging preclinical and clinical research.
- Radiomics outperforms traditional metrics: The developed RadSig consistently outperformed standard SUV metrics (SUVmean, SUVmax, SULpeak) in both predicting and assessing chemotherapy response in TNBC, both in preclinical models and in the clinical trial.
- Feature robustness is critical: Rigorous evaluation of radiomic features for reproducibility, cross-correlation, and volume dependency is essential for building reliable predictive models. Features that are not robust can introduce noise and bias.
- Predicting response is more feasible than assessing it: The study found that RadSig generally had a greater impact on predicting response (using baseline features) compared to assessing response (using changes in features).
- TNBC heterogeneity is key: The paper acknowledges that TNBC's inherent heterogeneity is a major factor in treatment resistance, and radiomics offers a way to capture this heterogeneity quantitatively.
Key Results
- Robust Features: Out of 131 initial features, 64 were identified as robust.
- Preclinical Performance: In the preclinical PDX trial, Naïve Bayes-RadSig achieved the highest accuracy (86.21% for prediction, 86.9% for assessment) using only 4 features. RadSig improved prediction by over 60% compared to SUV metrics.
- Clinical Performance: In the interim clinical analysis, SVM-RadSig and NB-RadSig showed practically identical performance, achieving accuracies of around 72% for prediction and 71% for assessment. This was superior to SUV metrics.
- Feature Reduction: The study successfully reduced the number of features to a concise "radiomic signature" (RadSig) of just 4 features for both prediction and assessment tasks, making it more practical for clinical implementation.
Limitations and Future Work
- Small Clinical Sample Size: The clinical validation was based on an interim analysis of only 20 patients. The authors emphasize the need for additional studies with larger cohorts to validate the utility of PDX-optimized features.
- Integration of Multi-scale Features: The authors suggest that integrating multi-scale features, including MR imaging features, could further enhance prediction and assessment capabilities.
- MR Radiomic Features: MR radiomic features were not included in this study due to increased dimensionality and a limited number of patients, but they are identified as a potential area for future integration.
- Other Technologies: The paper notes that other technologies like liquid biopsies (ctDNA) and molecular/genomic features could be integrated to further improve therapeutic prediction.
- Standardization: A critical aspect for the practical implementation of radiomics is the need for consensus on the most effective features and their standardization.
Summary updated