Development of a Radiomic Signature from Pre-Treatment FDG-PET for Recurrence Prediction in Cervical Cancer.
Int J Radiat Oncol Biol Phys 2023;
117:e469. [PMID:
37785492 DOI:
10.1016/j.ijrobp.2023.06.1674]
[Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [Track Full Text] [Journal Information] [Subscribe] [Scholar Register] [Indexed: 10/04/2023]
Abstract
PURPOSE/OBJECTIVE(S)
As many as 30-50% of patients with locally advanced cervical cancer (LACC) experience recurrence after standard-of-care chemoradiation therapy (CRT), creating a critical need to identify pre-treatment biomarkers of treatment failure. The purpose of this study is to identify whether radiomic features derived from pre-treatment FDG-PET imaging can be used to construct a predictive signature useful for assessing risk of recurrence during treatment planning.
MATERIALS/METHODS
Standardized uptake values (SUV) were obtained from within the physician-defined metabolic tumor volumes (MTV) delineated on FDG-PET scans acquired for 90 LACC patients at our institution prior to standard of care curative-intent CRT. Clinical outcome data of these patients has a median follow-up time of 85 months. The clinical endpoint was local recurrence within 3 years of treatment. 851 quantitative radiomic features describing intensity, shape, texture and high and low frequency spatial filters of the MTV were extracted for each patient. Low information features, defined by pairwise correlation > 0.85 with another feature or a maximum deviation within 20% of the mean, were discarded, leaving 146 features. Predictive signatures were constructed from features using multiple techniques, including multivariate Cox modelling, a set of machine learning models (random forest (RF), support vector classifier (SVC), ridge regression, LASSO regression, and elastic net regression), and a deep neural network (DNN) classifier.
RESULTS
The DNN classifier had the best overall performance, predicting a patient's recurrence group with an F1 score of 0.917 ± 0.028 under 5-fold cross-validation. By contrast, the Cox model classifier yielded an F1 score of 0.604 ± 0.085 and the best performing of the alternative machine learning models, elastic net, yielded F1 score of 0.868 ± 0.018. A set of textural features contributed the most to the output of the DNN classifier, including Large Area Low Gray Level Emphasis of the gray level size zone matrix (GLSZM) and coarseness and busyness of the neighboring gray tone difference matrix (NGTDM), reflecting the important role patterns of tumor heterogeneity play in post-treatment recurrence.
CONCLUSION
In this pilot study, we investigated multiple techniques to construct predictive radiomic signatures for local recurrence in LACC, determining that a DNN classifier is most capable of stratifying patients by risk of early recurrence. Future work will seek to validate this result on additional PET imaging data sets and to integrate radiomic features with gene expression data from matched tumor samples to establish radiogenomic biomarkers for recurrence.
Collapse