Radiomics-based machine-learning approach to predict response at brachytherapy using pretreatment magnetic resonance imaging in locally advanced cervical cancer.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiation Oncology, Tata Memorial Centre, Parel, Mumbai, Maharashtra, India; Homi Bhabha National Institute, Anushaktinagar, Trombay, Mumbai, Maharashtra, India.
- Department of Radiation Oncology, Tata Memorial Centre, Parel, Mumbai, Maharashtra, India; Homi Bhabha National Institute, Anushaktinagar, Trombay, Mumbai, Maharashtra, India. Electronic address: [email protected].
- Epidemological and Clinical Trials Unit, Advance Centre for Treatment Research Education in Cancer (ACTREC) Tata Memorial Centre, Kharghar, Mumbai, Maharashtra, India; Homi Bhabha National Institute, Anushaktinagar, Trombay, Mumbai, Maharashtra, India.
- Department of Radiodiagnosis, Tata Memorial Centre, Parel, Mumbai, Maharashtra, India; Homi Bhabha National Institute, Anushaktinagar, Trombay, Mumbai, Maharashtra, India.
- Department of Medical Oncology, Tata Memorial Centre, Parel, Mumbai, Maharashtra, India; Homi Bhabha National Institute, Anushaktinagar, Trombay, Mumbai, Maharashtra, India.
Abstract
We investigated baseline magnetic resonance imaging (MRI) radiomic features for predicting tumor response in patients with locally advanced cervical cancer (LACC) at brachytherapy (BT). Seventy-four patients underwent baseline T2W MRI. Gross tumor volume at diagnosis (GTV-T initial) was delineated. Tumor radiomic features were extracted using TexRAD software. Feature enrichment using parameters indicative of response was done using least absolute shrinkage and selection operator (LASSO) regression. The support vector machine (SVM) algorithm was used to generate the model. Response to chemo-radiotherapy was based on the criteria GTV-BT/GTV-T initial ratio < or >0.20 was used for classifying good versus poor responders. Fifty-six radiomic features were extracted. LASSO enriched the number of features to 11 for the GTV-BT/GTV-T initial ratio. The SVM classifier with a 10-fold internal cross-validation demonstrated an AUC of 0.82 and 76.8% accuracy when the response was assessed using the GTV-BT/GTV-T initial ratio for response evaluation. When SVM was modeled using clinical features, the AUC was 0.55, and the accuracy was 62.6% for the GTV-BT/GTV-T initial ratio, CONCLUSION: Machine learning model employing radiomic features extracted from pre-treatment MRI reliably predicted treatment response in patients with LACC.