A general framework to develop a radiomic fingerprint for progression-free survival in cervical cancer.

PURPOSE: Treatment of locally advanced cervical cancer patients includes chemoradiation followed by brachytherapy. Our aim is to develop a delta radiomics (DRF) model from MRI-based brachytherapy treatment and assess its association with progression free survival (PFS). MATERIALS AND METHODS: A retrospective analysis of FIGO stage IB- IV cervical cancer patients between 2012 and 2018 who were treated with definitive chemoradiation followed by MRI-based intracavitary brachytherapy was performed. Clinical factors together with 18 radiomic features extracted from different radiomics matrices were analyzed. The delta radiomic features (DRFs) were extracted from MRI on the first and last brachytherapy fractions. Support Vector Machine (SVM) models were fitted to combinations of 2-3 DRFs found significant after Spearman correlation and Wilcoxon rank sum test statistics. Additional models were tested that included clinical factors together with DRFs. RESULTS: A total of 39 patients were included in the analysis with a median patient age of 52 years. Progression occurred in 20% of patients (8/39). The significant DRFs using two DRF feature combinations was a model using auto correlation (AC) and sum variance (SV). The best performing three feature model combined mean, AC & SV. Additionally, the inclusion of FIGO stages with the 2- and 3 DRF combination model(s) improved performance compared to models with only DRFs. However, all the clinical factor + DRF models were not significantly different from one another (all AUCs were 0.77). CONCLUSIONS: Our study shows promising evidence that radiomics metrics are associated with progression free survival in cervical cancer.

Auto-detection of necessity for MRI-guided online adaptive replanning using a machine learning classifier.

PURPOSE: MRI-guided adaptive radiation therapy (MRgART), particularly daily online adaptive replanning (OLAR) can substantially improve radiation therapy delivery, however, it can be labor-intensive and time-consuming. Currently, the decision to perform OLAR for a treatment fraction is determined subjectively. In this work, we develop a machine learning algorithm based on structural similarity index measure (SSIM) and change in entropy to quickly and objectively determine whether OLAR is necessary for a daily MRI set. METHODS: A total of 109 daily MRI sets acquired on a 1.5T MR-Linac during MRgART for 22 pancreatic cancer patients each treated with five fractions were retrospectively analyzed. For each daily MRI set, OLAR and reposition (No-OLAR) plans were created and the superior plan with the daily fraction determined per clinical dose-volume criteria. SSIM and entropy maps were extracted from each daily MRI set, with respect to its reference (e.g., dry-run) MRI in the region enclosed by 50-100% isodose surfaces. A total of six common features were extracted from SSIM maps. Pearson's rank correlation coefficient was utilized to rule out redundant SSIM features. A t-test was used to determine significant SSIM features which were combined with the change in entropy to develop anensemble machine classifier with fivefold cross validation. The performance of the classifier was evaluated using the area under the curve (AUC) of the receiver operating characteristic curve. RESULTS: A machine learning classifier model using two SSIM features (mean and full width at half maximum) and change in entropy was determined to be able to significantly discriminate between No-OLAR and OLAR groups. The obtained machine learning ensemble classifier can predict OLAR necessity with a cross validated AUC of 0.93. Misclassification was found primarily for No-OLAR cases with dosimetric plan quality closely comparable to the corresponding OLAR plans, thus, are not a major practical concern. CONCLUSION: A machine learning classifier based on simple first-order image features, that is, SSIM features and change in entropy, was developed to determine when OLAR is necessary for a daily MRI set with practical acceptable prediction accuracy. This classifier may be implemented in the MRgART process to automatically and objectively determine if OLAR is required following daily MRI.

Predicting necessity of daily online adaptive replanning based on wavelet image features for MRI guided adaptive radiation therapy.

PURPOSE: Online adaptive replanning (OLAR) is generally labor-intensive and time-consuming during MRI-guided adaptive radiation therapy (MRgART). This work aims to develop a method to determine OLAR necessity during MRgART. METHODS: A machine learning classifier was developed to predict OLAR necessity based on wavelet multiscale texture features extracted from daily MRIs and was trained and tested with data from 119 daily MRI datasets acquired during MRgART for 24 pancreatic cancer patients treated on a 1.5 T MR-Linac. Spearman correlations, interclass correlation (ICC), coefficient of variance (COV), t-test (p < 0.05), self-organized map (SOM) and maximum stable extremal region (MSER) algorithm were used to determine candidate features, which were used to build the prediction models using Bayesian classifiers. The model performance was judged using the AUC of the ROC curve. RESULTS: Spearman correlation identified 123 features that were not redundant (r < 0.9). Of them 82 showed high ICC for repositioning > 0.6, 67 had a COV greater than 9% for OLAR. Among the 38 features passed the t-test, 25 passed the SOM and 12 passed the MSER. These final 12 features were used to build the classifier model. The combination of 2-3 features at a time was used to build the classifier models. The best performing model was a 3-feature combination, which can predict OLAR necessity with a CV-AUC of 0.98. CONCLUSIONS: A machine learning classifier model based on the wavelet features extracted from daily MRI for pancreatic cancer was developed to automatically and objectively determine if OLAR is necessary for a treatment fraction avoiding unnecessary effort during MRgART.

Auto-segmentation of pancreatic tumor in multi-parametric MRI using deep convolutional neural networks.

PURPOSE: The recently introduced MR-Linac enables MRI-guided Online Adaptive Radiation Therapy (MRgOART) of pancreatic cancer, for which fast and accurate segmentation of the gross tumor volume (GTV) is essential. This work aims to develop an algorithm allowing automatic segmentation of the pancreatic GTV based on multi-parametric MRI using deep neural networks. METHODS: We employed a square-window based convolutional neural network (CNN) architecture with three convolutional layer blocks. The model was trained using about 245,000 normal and 230,000 tumor patches extracted from 37 DCE MRI sets acquired in 27 patients with data augmentation. These images were bias corrected, intensity standardized, and resampled to a fixed voxel size of 1 × 1 × 3 mm3. The trained model was tested on 19 DCE MRI sets from another 13 patients, and the model-generated GTVs were compared with the manually segmented GTVs by experienced radiologist and radiation oncologists based on Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and Mean Surface Distance (MSD). RESULTS: The mean values and standard deviations of the performance metrics on the test set were DSC = 0.73 ± 0.09, HD = 8.11 ± 4.09 mm, and MSD = 1.82 ± 0.84 mm. The interobserver variations were estimated to be DSC = 0.71 ± 0.08, HD = 7.36 ± 2.72 mm, and MSD = 1.78 ± 0.66 mm, which had no significant difference with model performance at p values of 0.6, 0.52, and 0.88, respectively. CONCLUSION: We developed a CNN-based model for auto-segmentation of pancreatic GTV in multi-parametric MRI. Model performance was comparable to expert radiation oncologists. This model provides a framework to incorporate multimodality images and daily MRI for GTV auto-segmentation in MRgOART.

Indications of Online Adaptive Replanning Based On Organ Deformation.

PURPOSE: Although vital to account for interfractional variations during radiation therapy, online adaptive replanning (OLAR) is time-consuming and labor-intensive compared with the repositioning method. Repositioning is enough for minimal interfractional deformations. Therefore, determining indications for OLAR is desirable. We introduce a method to rapidly determine the need for OLAR by analyzing the Jacobian determinant histogram (JDH) obtained from deformable image registration between reference (planning) and daily images. METHODS AND MATERIALS: The proposed method was developed and tested based on daily computed tomography (CT) scans acquired during image guided radiation therapy for prostate cancer using an in-room CT scanner. Deformable image registration between daily and reference CT scans was performed. JDHs were extracted from the prostate and a uniform surrounding 10-mm expansion. A classification tree was trained to determine JDH metrics to predict the need for OLAR for a daily CT set. Sixty daily CT scans from 12 randomly selected prostate cases were used as the training data set, with dosimetric plans for both OLAR and repositioning used to determine their class. The resulting classification tree was tested using an independent data set of 45 daily CT scans from 9 other patients with 5 CT scans each. RESULTS: Of a total of 27 JDH metrics tested, 5 were identified predicted whether OLAR was substantially superior to repositioning for a given fraction. A decision tree was constructed using the obtained metrics from the training set. This tree correctly identified all cases in the test set where benefits of OLAR were obvious. CONCLUSIONS: A decision tree based on JDH metrics to quickly determine the necessity of online replanning based on the image of the day without segmentation was determined using a machine learning process. The process can be automated and completed within a minute, allowing users to quickly decide which fractions require OLAR.

Automated air region delineation on MRI for synthetic CT creation.

Automatically and accurately separating air from other low signal regions (especially bone, liver, etc) in an MRI is difficult because these tissues produce similar MR intensities, resulting in errors in synthetic CT generation for MRI-based radiation therapy planning. This work aims to develop a technique to accurately and automatically determine air-regions for MR-guided adaptive radiation therapy. CT and MRI scans (T2-weighted) of phantoms with fabricated air-cavities and abdominal cancer patients were used to establish an MR intensity threshold for air delineation. From the phantom data, air/tissue boundaries in MRI were identified by CT-MRI registration. A formula relating the MRI intensities of air and surrounding materials was established to auto-threshold air-regions. The air-regions were further refined by using quantitative image texture features. A naive Bayesian classifier was trained using the extracted features with a leave-one-out cross validation technique to differentiate air from non-air voxels. The multi-step air auto-segmentation method was tested against the manually segmented air-regions. The dosimetry impacts of the air-segmentation methods were studied. Air-regions in the abdomen can be segmented on MRI within 1 mm accuracy using a multi-step auto-segmentation method as compared to manually delineated contours. The air delineation based on the MR threshold formula was improved using the MRI texture differences between air and tissues, as judged by the area under the receiver operating characteristic curve of 81% when two texture features (autocorrelation and contrast) were used. The performance increased to 82% with using three features (autocorrelation, sum-variance, and contrast). Dosimetric analysis showed no significant difference between the auto-segmentation and manual MR air delineation on commonly used dose volume parameters. The proposed techniques consisting of intensity-based auto-thresholding and image texture-based voxel classification can automatically and accurately segment air-regions on MRI, allowing synthetic CT to be generated quickly and precisely for MR-guided adaptive radiation therapy.

A machine learning based delta-radiomics process for early prediction of treatment response of pancreatic cancer.

Changes of radiomic features over time in longitudinal images, delta radiomics, can potentially be used as a biomarker to predict treatment response. This study aims to develop a delta-radiomic process based on machine learning by (1) acquiring and registering longitudinal images, (2) segmenting and populating regions of interest (ROIs), (3) extracting radiomic features and calculating their changes (delta-radiomic features, DRFs), (4) reducing feature space and determining candidate DRFs showing treatment-induced changes, and (5) creating outcome prediction models using machine learning. This process was demonstrated by retrospectively analyzing daily non-contrast CTs acquired during routine CT-guided-chemoradiation therapy for 90 pancreatic cancer patients. A total of 2520 CT sets (28-daily-fractions-per-patient) along with their pathological response were analyzed. Over 1300 radiomic features were extracted from the segmented ROIs. Highly correlated DRFs were ruled out using Spearman correlations. Correlation between the selected DRFs and pathological response was established using linear-regression-models. T test and linear-mixed-effects-models were used to determine which DRFs changed significantly compared with first fraction. A Bayesian-regularization-neural-network was used to build a response prediction model. The model was trained using 50 patients and leave-one-out-cross-validation. Performance was judged using the area-under-ROC-curve. External independent validation was done using data from the remaining 40 patients. The results show that 13 DRFs passed the tests and demonstrated significant changes following 2-4 weeks of treatment. The best performing combination differentiating good versus bad responders (CV-AUC = 0.94) was obtained using normalized-entropy-to-standard-deviation-difference-(NESTD), kurtosis, and coarseness. With further studies using larger data sets, delta radiomics may develop into a biomarker for early prediction of treatment response.

A Quantitative Ultrasound-Based Multi-Parameter Classifier for Breast Masses.

This manuscript reports preliminary results obtained by combining estimates of two or three (among seven) quantitative ultrasound (QUS) parameters in a model-free, multi-parameter classifier to differentiate breast carcinomas from fibroadenomas (the most common benign solid tumor). Forty-three patients scheduled for core biopsy of a suspicious breast mass were recruited. Radiofrequency echo signal data were acquired using clinical breast ultrasound systems equipped with linear array transducers. The reference phantom method was used to obtain system-independent estimates of the specific attenuation (ATT), the average backscatter coefficients, the effective scatterer diameter (ESD) and an effective scatterer diameter heterogeneity index (ESDHI) over regions of interest within each mass. In addition, the envelope amplitude signal-to-noise ratio (SNR), the Nakagami shape parameter, m, and the maximum collapsed average (maxCA) of the generalized spectrum were also computed. Classification was performed using the minimum Mahalanobis distance to the centroids of the training classes and tested against biopsy results. Classification performance was evaluated with the area under the receiver operating characteristic (ROC) curve. The best performance with a two-parameter classifier used the ESD and ESDHI and resulted in an area under the ROC curve of 0.98 (95% confidence interval [CI]: 0.95-1.00). Classification performance improved with three parameters (ATT, ESD and ESDHI) yielding an area under the ROC curve of 0.999 (0.995-1.000). These results suggest that system-independent QUS parameters, when combined in a model-free classifier, are a promising tool to characterize breast tumors. A larger study is needed to further test this idea.

Improving Treatment Response Prediction for Chemoradiation Therapy of Pancreatic Cancer Using a Combination of Delta-Radiomics and the Clinical Biomarker CA19-9.

Recently we showed that delta radiomics features (DRF) from daily CT-guided chemoradiation therapy (CRT) is associated with early prediction of treatment response for pancreatic cancer. CA19-9 is a widely used clinical biomarker for pancreatic cancer. The purpose of this work is to investigate if the predictive power of such biomarkers (DRF or CA19-9) can improve by combining both biomarkers. Daily non-contrast CTs acquired during routine CT-guided neoadjuvant CRT for 24 patients (672 datasets, in 28 daily fractions), along with their CA19-9, pathology reports and follow-up data were analyzed. The pancreatic head was segmented on each daily CT and radiomic features were extracted from the segmented regions. The time between the end of treatment and last follow-up was used to build a survival model. Patients were divided into two groups based on their pathological response. Spearman correlations were used to find the DRFs correlated to CA19-9. A regression model was built to examine the effect of combining CA19-9 and DRFs on response prediction. C-index was used to measure model effectiveness. The effect of a decrease in CA19-9 levels during treatment vs. failure of CA19-9 levels to normalize on survival was examined. Univariate- and multivariate Cox-regression analysis were performed to determine the effect of combining CA19-9 and DRFs on survival correlations. Spearman correlation showed that CA19-9 is correlated to DRFs (Entropy, cluster tendency and coarseness). An Increase in CA19-9 levels during treatment were correlated to a bad response, while a decline was correlated to a good response. Incorporating CA19-9 with DRFs increased the c-index from 0.57 to 0.87 indicating a stronger model. The univariate analysis showed that patients with decreasing CA19-9 had an improved median survival (68 months) compared to those with increasing levels (33 months). The 5-years survival was improved for the decreasing CA19-9 group (55%) compared to the increasing group (30%). The Cox-multivariate analysis showed that treatment related decrease in CA19-9 levels (p = 0.031) and DRFs (p = 0.001) were predictors of survival. The hazard-ratio was reduced from 0.73, p = 0.032 using CA19-9 only to 0.58, p = 0.028 combining DRFs with CA19-9. DRFs-CA19-9 combination has the potential to increasing the possibility for response-based treatment adaptation.

Three-dimensional Ultrasound Elasticity Imaging on an Automated Breast Volume Scanning System.

Ultrasound elasticity imaging has demonstrated utility in breast imaging, but it is typically performed with handheld transducers and two-dimensional imaging. Two-dimensional (2D) elastography images tissue stiffness of only a plane and hence suffers from errors due to out-of-plane motion, whereas three-dimensional (3D) data acquisition and motion tracking can be used to track out-of-plane motion that is lost in 2D elastography systems. A commercially available automated breast volume scanning system that acquires 3D ultrasound data with precisely controlled elevational movement of the 1D array ultrasound transducer was employed in this study. A hybrid guided 3D motion-tracking algorithm was developed that first estimated the displacements in one plane using a modified quality-guided search method, and then performed an elevational guided-search for displacement estimation in adjacent planes. To assess the performance of the method, 3D radiofrequency echo data were acquired with this system from a phantom and from an in vivo human breast. For both experiments, the axial displacement fields were smooth and high cross-correlation coefficients were obtained in most of the tracking region. The motion-tracking performance of the new method was compared with a correlation-based exhaustive-search method. For all motion-tracking volume pairs, the average motion-compensated cross-correlation values obtained by the guided-search motion-tracking method were equivalent to those by the exhaustive-search method, and the computation time was about a factor of 10 lesser. Therefore, the proposed 3D ultrasound elasticity imaging method was a more efficient approach to produce a high quality of 3D ultrasound strain image.

Acoustic Properties of Breast Fat.

OBJECTIVES: The American College of Radiology Breast Imaging Reporting and Data System (BI-RADS) atlas for ultrasound (US) qualitatively describes the echogenicity and attenuation of a mass, where fat lobules serve as a standard for comparison. This study aimed to estimate acoustic properties of breast fat under clinical imaging conditions to determine the degree to which properties vary among patients. METHODS: Twenty-four women with solid breast masses scheduled for biopsy were scanned with a Siemens S2000 scanner and 18L6 linear array transducer (Siemens Medical Solutions, Malvern, PA). Offline analysis estimated the attenuation coefficient and backscatter coefficients (BSCs) from breast fat using the reference phantom method. The average BSC was calculated over 6 to 12 MHz to objectively quantify the BI-RADS US echo pattern descriptor, and effective scatterer diameters were also estimated. RESULTS: A power law fit to the attenuation coefficient versus frequency yielded an attenuation coefficient of 1.28 dB·cm(-1) MHz(-0.73). The mean attenuation coefficient versus frequency slope ± SD at 7 MHz was 0.73 ± 0.23 dB·cm(-1) MHz(-1), in agreement with previously reported values. The BSC versus frequency showed close agreement among all patients, both in magnitude and frequency dependence, with a power law fit of (0.6 ± 0.25) ×10(-4) sr(-1) cm(-1) MHz(-2.49). The average backscatter in the 6-12-MHz range was 0.004 ± 0.002 sr(-1) cm(-1). The mean effective scatterer diameter for fat was 60.2 ± 9.5 µm. CONCLUSIONS: The agreement in parameter estimates for breast fat among these patients supports the use of fat as a standard for comparison with tumors. Results also suggest that objective quantification of these BI-RADS US descriptors may reduce subjectivity when interpreting B-mode image data.