Privacy-preserving distributed learning of radiomics to predict overall survival and HPV status in head and neck cancer

Abstract

A major challenge in radiomics is assembling data from multiple centers. Sharing data between hospitals is restricted by legal and ethical regulations. Distributed learning is a technique, enabling training models on multicenter data without data leaving the hospitals ("privacy-preserving" distributed learning). This study tested feasibility of distributed learning of radiomics data for prediction of two year overall survival and HPV status in head and neck cancer (HNC) patients. Pretreatment CT images were collected from 1174 HNC patients in 6 different cohorts. 981 radiomic features were extracted using Z-Rad software implementation. Hierarchical clustering was performed to preselect features. Classification was done using logistic regression. In the validation dataset, the receiver operating characteristics (Roc) were compared between the models trained in the centralized and distributed manner. No difference in ROC was observed with respect to feature selection. The logistic regression coefficients were identical between the methods (absolute difference <10 −7). In comparison of the full workflow (feature selection and classification), no significant difference in ROC was found between centralized and distributed models for both studied endpoints (DeLong p > 0.05). In conclusion, both feature selection and classification are feasible in a distributed manner using radiomics data, which opens new possibility for training more reliable radiomics models.

Application of Radiomics for the Prediction of Radiation-Induced Toxicity in the IMRT Era: Current State-of-the-Art

“…In the perspective of personalized oncology ( 13 ) as currently implemented in the clinic, the use of quantitative imaging may allow us to overcome the known limits associated with molecular profiling. Several applications of radiomics in the field of precision radiation oncology have been identified, providing insights in terms of stage discrimination ( 14 , 15 ), molecular stratification ( 16 – 18 ), prognostic impact ( 19 , 20 ), and prediction of response to treatment ( 21 – 23 ). With imaging, the possibility to capture intrinsic tumor and organ-specific heterogeneity could be leveraged to evaluate the individual predisposition to radiation-induced toxicity ( 24 ).…”

Imaging biomarkers of contrast-enhanced computed tomography predict survival in oesophageal cancer after definitive concurrent chemoradiotherapy

“…For geometric IBMs, sphericity, volume-density and the major axis length quantified the sphericity and size of tumours. Previous studies revealed that these IBMs basically represented tumour volume, which were significantly associated with treatment outcomes [ 30 , 31 ]. In our study, the discrimination performances of the IBM nomograms were decreased when volume-related IBMs were omitted from the IBM score (C-index for the radiomics nomogram: OS, 0.672 (95% CI, 0.588–0.757); PFS, 0.629 (95% CI, 0.545–0.713) in the training cohort).…”

“…For geometric IBMs, sphericity, volume-density and the major axis length quantified the sphericity and size of tumours. Previous studies revealed that these IBMs basically represented tumour volume, which were significantly associated with treatment outcomes [30,31]. Decision curve analysis of PFS and OS were compared between the IBM score and clinical stage in the training and validation cohort, respectively.…”

Imaging biomarkers of contrast-enhanced computed tomography predict survival in oesophageal cancer after definitive concurrent chemoradiotherapy

“…For geometric IBMs, sphericity, volume-density and the major axis length quanti ed the sphericity and size of tumours. Previous studies revealed that these IBMs basically represented tumour volume, which were signi cantly associated with treatment outcomes [30,31]. In our study, the discrimination performances of the IBM nomograms were decreased when volume-related IBMs were omitted from the IBM score (C-index for the radiomics nomogram: OS, 0.672 (95% CI, 0.588-0.757); PFS, 0.629 (95% CI, 0.545-0.713) in the training cohort).…”