AFFILIAZIONE
AO Regina Margherita- Sant’anna
AUTORE PRINCIPALE
Dr. Papa Francesco
Registrazione obbligatoria. Una valutazione consentita
login avvenuto
GRUPPO DI LAVORO
AREA TEMATICA
Applicazioni di intelligenza artificiale in sanità
ABSTRACT
Purpose or Learning Objective
Synovial sarcoma is an aggressive malignancy that necessitates innovative diagnostic and therapeutic strategies. Predicting chemotherapy response is a pivotal step in the diagnostic-therapeutic pathway, enabling personalized treatment approaches. Radiomics, combined with machine learning, offers a promising avenue to identify imaging biomarkers that predict treatment response.
Methods or Background
In this retrospective, monocentric study, 51 patients diagnosed with synovial sarcoma underwent peri-treatment MRI acquisition. T1-weighted sequences were analyzed to extract 1116 radiomic features, which were subjected to feature selection processes. Machine learning models were developed to predict chemotherapy response. Data were intentionally gathered from heterogeneous MRI scanners to ensure the robustness of the models against clinical variability. Performance metrics included area under the curve (AUC) and accuracy assessments on both training and validation sets.
Results or Findings
The final ExtraTrees model, trained on 20 selected features, showed promising performance on the training set with an accuracy of 0.81 (±0.34), precision of 0.88 (±0.37), recall of 0.87 (±0.42), F1-score of 0.84 (±0.28), and AUC of 0.89 (±0.35), despite high variability.
On the test set (n=17), accuracy was 64.7%, with good sensitivity for non-responders (80%) but lower for responders (58%). Precision was high for responders (0.88) but lower for non-responders (0.44). The test AUC was modest (0.60), while the Precision-Recall Curve (0.72) indicated reasonable precision for the positive class. The Brier score (0.250) suggested moderate model calibration.
Conclusion
This study highlights the potential of radiomics to predict chemotherapy response in synovial sarcoma, despite challenges posed by scanner heterogeneity and intra-reader variability. Further research is required to refine these methods and integrate them into routine clinical practice.
Limitations
This study is limited by single-reader segmentation, small sample size, and lack of external validation.
Synovial sarcoma is an aggressive malignancy that necessitates innovative diagnostic and therapeutic strategies. Predicting chemotherapy response is a pivotal step in the diagnostic-therapeutic pathway, enabling personalized treatment approaches. Radiomics, combined with machine learning, offers a promising avenue to identify imaging biomarkers that predict treatment response.
Methods or Background
In this retrospective, monocentric study, 51 patients diagnosed with synovial sarcoma underwent peri-treatment MRI acquisition. T1-weighted sequences were analyzed to extract 1116 radiomic features, which were subjected to feature selection processes. Machine learning models were developed to predict chemotherapy response. Data were intentionally gathered from heterogeneous MRI scanners to ensure the robustness of the models against clinical variability. Performance metrics included area under the curve (AUC) and accuracy assessments on both training and validation sets.
Results or Findings
The final ExtraTrees model, trained on 20 selected features, showed promising performance on the training set with an accuracy of 0.81 (±0.34), precision of 0.88 (±0.37), recall of 0.87 (±0.42), F1-score of 0.84 (±0.28), and AUC of 0.89 (±0.35), despite high variability.
On the test set (n=17), accuracy was 64.7%, with good sensitivity for non-responders (80%) but lower for responders (58%). Precision was high for responders (0.88) but lower for non-responders (0.44). The test AUC was modest (0.60), while the Precision-Recall Curve (0.72) indicated reasonable precision for the positive class. The Brier score (0.250) suggested moderate model calibration.
Conclusion
This study highlights the potential of radiomics to predict chemotherapy response in synovial sarcoma, despite challenges posed by scanner heterogeneity and intra-reader variability. Further research is required to refine these methods and integrate them into routine clinical practice.
Limitations
This study is limited by single-reader segmentation, small sample size, and lack of external validation.
