M. Vallieres

18F-FDG PET Uptake Characterization Through Texture Analysis: Investigating the Complementary Nature of Heterogeneity and Functional Tumor Volume in a Multi–Cancer Site Patient Cohort

Intratumoral uptake heterogeneity in 18F-FDG PET has been associated with patient treatment outcomes in several cancer types. Textural feature analysis is a promising method for its quantification. An open issue associated with textural features for the quantification of intratumoral heterogeneity concerns its added contribution and dependence on the metabolically active tumor volume (MATV), which has already been shown to be a significant predictive and prognostic parameter. Our objective was to address this question using a larger cohort of patients covering different cancer types. Methods: A single database of 555 pretreatment 18F-FDG PET images (breast, cervix, esophageal, head and neck, and lung cancer tumors) was assembled. Four robust and reproducible textural feature–derived parameters were considered. The issues associated with the calculation of textural features using co-occurrence matrices (such as the quantization and spatial directionality relationships) were also investigated. The relationship between these features and MATV, as well as among the features themselves, was investigated using Spearman rank coefficients for different volume ranges. The complementary prognostic value of MATV and textural features was assessed through multivariate Cox analysis in the esophageal and non–small cell lung cancer (NSCLC) cohorts. Results: A large range of MATVs was included in the population considered (3–415 cm3; mean, 35; median, 19; SD, 50). The correlation between MATV and textural features varied greatly depending on the MATVs, with reduced correlation for increasing volumes. These findings were reproducible across the different cancer types. The quantization and calculation methods both had an impact on the correlation. Volume and heterogeneity were independent prognostic factors (P = 0.0053 and 0.0093, respectively) along with stage (P = 0.002) in non–small cell lung cancer, but in the esophageal tumors, volume and heterogeneity had less complementary value because of smaller overall volumes. Conclusion: Our results suggest that heterogeneity quantification and volume may provide valuable complementary information for volumes above 10 cm3, although the complementary information increases substantially with larger volumes.

A radiomics model from joint FDG-PET and MRI texture features for the prediction of lung metastases in soft-tissue sarcomas of the extremities

This study aims at developing a joint FDG-PET and MRI texture-based model for the early evaluation of lung metastasis risk in soft-tissue sarcomas (STSs). We investigate if the creation of new composite textures from the combination of FDG-PET and MR imaging information could better identify aggressive tumours. Towards this goal, a cohort of 51 patients with histologically proven STSs of the extremities was retrospectively evaluated. All patients had pre-treatment FDG-PET and MRI scans comprised of T1-weighted and T2-weighted fat-suppression sequences (T2FS). Nine non-texture features (SUV metrics and shape features) and forty-one texture features were extracted from the tumour region of separate (FDG-PET, T1 and T2FS) and fused (FDG-PET/T1 and FDG-PET/T2FS) scans. Volume fusion of the FDG-PET and MRI scans was implemented using the wavelet transform. The influence of six different extraction parameters on the predictive value of textures was investigated. The incorporation of features into multivariable models was performed using logistic regression. The multivariable modeling strategy involved imbalance-adjusted bootstrap resampling in the following four steps leading to final prediction model construction: (1) feature set reduction; (2) feature selection; (3) prediction performance estimation; and (4) computation of model coefficients. Univariate analysis showed that the isotropic voxel size at which texture features were extracted had the most impact on predictive value. In multivariable analysis, texture features extracted from fused scans significantly outperformed those from separate scans in terms of lung metastases prediction estimates. The best performance was obtained using a combination of four texture features extracted from FDG-PET/T1 and FDG-PET/T2FS scans. This model reached an area under the receiver-operating characteristic curve of 0.984 ± 0.002, a sensitivity of 0.955 ± 0.006, and a specificity of 0.926 ± 0.004 in bootstrapping evaluations. Ultimately, lung metastasis risk assessment at diagnosis of STSs could improve patient outcomes by allowing better treatment adaptation.

SU-E-J-250: A Machine Learning Approach for Creating Texture-Preserved MRI Tumor Models From Clinical Sequences

Purpose: We hypothesize that MRI texture-based tumor outcome prediction models could be optimized via numerical simulations of image acquisitions. These simulations require knowledge of T1 and T2 relaxation times as inputs. The goal of this study is to evaluate the feasibility of using machine learning techniques to infer T1 and T2 tumor maps with accurate texture preservation for simulation inputs from clinical sequences. Methods: Clinical T1-weighted (T1w) and T2-weighted fat-saturated (T2FS) scans, and measured T1 and T2 maps from eight patients with soft-tissue sarcomas were used in this study. Measured T1 and T2 maps were computed using pulse sequences with variable flip angles and echo times, respectively. General regression neural networks (GRNNs) were trained on these data to infer T1 and T2 relaxation times from T1w and T2FS images. Four texture features were extracted to evaluate texture preservation: GLCM/Entropy, GLRLM/Gray-Level Variance (GLV), GLSZM/Zone Size Variance (ZSV) and NGTDM/Complexity. The GRNN ability to estimate T1 and T2 relaxation times was assessed using leave-one-out cross-validation. Results: The average T1 and T2 relaxation times within the tumor region of all patients were (1515 ± 542) ms and (226 ± 151) ms in the measured cases, and (1546 ± 546) ms and (249 ± 145) ms in the estimated cases, respectively. The average root-mean-square errors between measured and estimated relaxation times were 573 ms for T1 and 160 ms for T2. The average absolute percentage differences between measured and estimated GLCM/Entropy, GLRLM/GLV, GLSZM/ZSV and NGTDM/Complexity features were 5.1%, 0.02%, 0.0% and 16.2% for T1 maps, and 7.7%, 0.04%, 0.0% and 10.9% for T2 maps, respectively. Conclusion: From a texture preservation perspective, this work demonstrates the feasibility to create MRI numerical models using GRNNs from T1w and T2FS clinical scans. Further work is required to obtain higher accuracy for T1 and T2 absolute relaxation times. This work was supported by the Natural Sciences and Engineering Research Council (NSERC) of Canada under the scholarship CGSD3-426742-2012, as well as it was supported by the Canadian Institutes of Health Research (CIHR) under grant MOP-136774.

WE‐C‐WAB‐02: Joint FDG‐PET/MR Imaging for the Early Prediction of Tumor Outcomes

PURPOSE To investigate the potential of joint FDG-PET/MR imaging features for the prediction of lung metastases at diagnosis of soft-tissue sarcomas (STS). METHODS A cohort of 35 patients with histologically proven STS was used in this study. All patients underwent pre-treatment FDG-PET and MR scans that comprised T1 and T2-fat suppression weighted (T2FS) sequences. The cohort had a median follow-up period of 29 months (range: 4-85) during which 13 patients developed lung metastases. An SUV feature (SUVmax) from the FDG-PET scans and 6 texture features (energy, entropy, contrast, homogeneity, sum-mean and variance) from the co-occurrence matrix of the separate (FDG-PET, T1 and T2FS) and fused (FDG-PET/T1 and FDG-PET/T2FS) scans were extracted from the tumor region. Fusion of the scans was implemented using the wavelet transform. Multivariable modeling was performed using logistic regression (LR) and the corresponding performance for lung metastases prediction was assessed using receiver operating characteristic (ROC) metrics on bootstrapping resampling. Optimal texture extraction was carried out through the optimization of intensity quantization, spatial resolution and wavelet band-pass filtering. RESULTS Overall, textures extracted from fused scans outperformed those from separate scans for the prediction of lung metastases. The best performance was found using an LR model with the following 4 parameters: SUVmax, FDG-PET/T1--contrast, FDG-PET/T1--homogeneity and FDG-PET/T2FS--variance. The average performance of this model in 10000 bootstrapping testing sets was: AUC=0.956, sensitivity=0.909, specificity=0.925, accuracy=0.916. However, the upper limit on the uncertainty of the texture model due to contouring variations was evaluated to be 15%. CONCLUSION Our results demonstrate that fused FDG-PET/MR texture features can be used to evaluate lung metastasis risk at diagnosis of STS. Accurate risk assessment could improve patient outcomes by allowing better adapted treatments. The methodology developed in this study could be tested on other cancer types and clinical endpoints such as treatment response.

TH‐A‐WAB‐02: FDG‐PET Imaging Features Can Predict Treatment Outcomes in Head and Neck Cancer

PURPOSE To investigate the potential of FDG-PET image-derived characteristics for the prediction of head neck cancer treatment outcomes. METHODS A cohort of 67 patients with histologically proven head and neck squamous cell carcinoma (HNSCC) was retrospectively evaluated in this study. All patients underwent pre-treatment FDG-PET scans before receiving radical radiotherapy (n=7) or chemo-radiotherapy (n=60). Patients had a median follow-up of 30 months (range: 4-71). Treatment failure (TF) was reported for 11 patients as tumor recurrence and/or distant metastases (DM, n=8). Eleven features were extracted from the FDG-PET tumor region: 6 texture features (energy, entropy, homogeneity, contrast, correlation and variance), 2 SUV measures (SUVmax and % inactive volume) and 3 shape features (volume, solidity and eccentricity). Multivariable modeling was performed using ensembles of logistic regression (LR) classifiers. The corresponding classification performance was assessed using receiver operating characteristic (ROC) metrics on leave-one-out cross-validation (LOO-CV) resampling. The LR ensembles accounted for the effect of data imbalance by repeating the TF/DM instances (n=11/8) into an optimal number M of partitions and by randomly distributing the non-TF/DM instances (n=56/59) into the M partitions (for 100 LOO-CV repetitions), to finally average the partitions LR responses. RESULTS The subset of features that yielded the highest area under the ROC curve (AUC) for TF prediction using M=7 was: entropy, variance, volume and solidity. This model reached an AUC of 0.73 (0.74 sensitivity, 0.63 specificity). Similarly, the prediction of DM with M=8 using an equivalent model (energy, variance, volume, solidity) reached an AUC of 0.77 (0.78 sensitivity, 0.67 specificity). CONCLUSION Our results demonstrate the possibility of using prognostic models combining tumor shape and FDG-PET texture features for the prediction of treatment outcomes in HNSCC. The ensemble methodology used in this study allowed the modeling of unbalanced data without compromising either the sensitivity or the specificity of the LR classifiers.

SU‐E‐J‐110: A Novel Level Set Active Contour Algorithm for Multimodality Joint Segmentation/Registration Using the Jensen‐Rényi Divergence

PURPOSE To present a novel joint segmentation/registration for multimodality image-guided and adaptive radiotherapy. A major challenge to this framework is the sensitivity of many segmentation or registration algorithms to noise. Presented is a level set active contour based on the Jensen-Renyi (JR) divergence to achieve improved noise robustness in a multi-modality imaging space. METHODS To present a novel joint segmentation/registration for multimodality image-guided and adaptive radiotherapy. A major challenge to this framework is the sensitivity of many segmentation or registration algorithms to noise. Presented is a level set active contour based on the Jensen-Renyi (JR) divergence to achieve improved noise robustness in a multi-modality imaging space. RESULTS It was found that JR divergence when used for segmentation has an improved robustness to noise compared to using mutual information, or other entropy-based metrics. The MI metric failed at around 2/3 the noise power than the JR divergence. CONCLUSIONS The JR divergence metric is useful for the task of joint segmentation/registration of multimodality images and shows improved results compared entropy based metric. The algorithm can be easily modified to incorporate non-intensity based images, which would allow applications into multi-modality and texture analysis.