Radiomics analysis of 3D dose distributions to predict toxicity of radiotherapy for lung cancer.

PURPOSE: (Chemo)-radiotherapy (RT) is the gold standard treatment for patients with locally advanced lung cancer non accessible for surgery. However, current toxicity prediction models rely on clinical and dose volume histograms (DVHs) and remain unsufficient. The goal of this work is to investigate the added predictive value of the radiomics approach applied to dose maps regarding acute and late toxicities in both the lungs and esophagus. METHODS: Acute and late toxicities scored using the CTCAE v4.0 were retrospectively collected on patients treated with RT in our institution. Radiomic features were extracted from 3D dose maps considering Gy values as grey-levels in images. DVH and usual clinical factors were also considered. Three toxicity prediction models (clinical only, clinical + DVH and combined, i.e., including clinical + DVH + radiomics) were incrementally trained using a neural network on 70% of the patients for prediction of grade ≥2 acute and late pulmonary toxicities (APT/LPT) and grade ≥2 acute esophageal toxicities (AET). After bootstrapping (n = 1000), optimal cut-off values were determined based on the Youden Index. The trained models were then evaluated in the remaining 30% of patients using balanced accuracy (BAcc). RESULTS: 167 patients were treated from 2015 to 2018: 78% non small-cell lung cancers, 14% small-cell lung cancers and 8% other histology with a median age at treatment of 66 years. Respectively, 22.2%, 16.8% and 30.0% experienced APT, LPT and AET. In the training set (n = 117), the corresponding BAcc for clinical only/clinical + DVH/combined were 0.68/0.79/0.92, 0.66/0.77/0.87 and 0.68/0.73/0.84. In the testing evaluation (n = 50), these trained models obtained a corresponding BAcc of 0.69/0.69/0.92, 0.76/0.80/0.89 and 0.58/0.73/0.72. CONCLUSION: In patients with a lung cancer treated with RT, radiomic features extracted from 3D dose maps seem to surpass usual models based on clinical factors and DVHs for the prediction of APT and LPT.

Harmonization strategies for multicenter radiomics investigations.

Carrying out large multicenter studies is one of the key goals to be achieved towards a faster transfer of the radiomics approach in the clinical setting. This requires large-scale radiomics data analysis, hence the need for integrating radiomic features extracted from images acquired in different centers. This is challenging as radiomic features exhibit variable sensitivity to differences in scanner model, acquisition protocols and reconstruction settings, which is similar to the so-called 'batch-effects' in genomics studies. In this review we discuss existing methods to perform data integration with the aid of reducing the unwanted variation associated with batch effects. We also discuss the future potential role of deep learning methods in providing solutions for addressing radiomic multicentre studies.

Performance comparison of modified ComBat for harmonization of radiomic features for multicenter studies.

Multicenter studies are needed to demonstrate the clinical potential value of radiomics as a prognostic tool. However, variability in scanner models, acquisition protocols and reconstruction settings are unavoidable and radiomic features are notoriously sensitive to these factors, which hinders pooling them in a statistical analysis. A statistical harmonization method called ComBat was developed to deal with the "batch effect" in gene expression microarray data and was used in radiomics studies to deal with the "center-effect". Our goal was to evaluate modifications in ComBat allowing for more flexibility in choosing a reference and improving robustness of the estimation. Two modified ComBat versions were evaluated: M-ComBat allows to transform all features distributions to a chosen reference, instead of the overall mean, providing more flexibility. B-ComBat adds bootstrap and Monte Carlo for improved robustness in the estimation. BM-ComBat combines both modifications. The four versions were compared regarding their ability to harmonize features in a multicenter context in two different clinical datasets. The first contains 119 locally advanced cervical cancer patients from 3 centers, with magnetic resonance imaging and positron emission tomography imaging. In that case ComBat was applied with 3 labels corresponding to each center. The second one contains 98 locally advanced laryngeal cancer patients from 5 centers with contrast-enhanced computed tomography. In that specific case, because imaging settings were highly heterogeneous even within each of the five centers, unsupervised clustering was used to determine two labels for applying ComBat. The impact of each harmonization was evaluated through three different machine learning pipelines for the modelling step in predicting the clinical outcomes, across two performance metrics (balanced accuracy and Matthews correlation coefficient). Before harmonization, almost all radiomic features had significantly different distributions between labels. These differences were successfully removed with all ComBat versions. The predictive ability of the radiomic models was always improved with harmonization and the improved ComBat provided the best results. This was observed consistently in both datasets, through all machine learning pipelines and performance metrics. The proposed modifications allow for more flexibility and robustness in the estimation. They also slightly but consistently improve the predictive power of resulting radiomic models.