Browsing by Subject "Radiomics"

Purpose: To compare radiomics and deep-learning (DL) methods for predicting NSCLC surgical treatment failure. Methods: A cohort of 83 patients undergoing lobectomy or wedge resection for early-stage NSCLC from our institution was studied. There were 7 local failures and 16 non-local failures (regional and/or distant). Gross tumor volumes (GTV) were contoured on pre-surgery CT datasets after 1mm3 isotropic resolution resampling. For the radiomics analysis, 92 radiomics features were extracted from the GTV and z-score normalizations were performed. The multivariate association between the extracted features and clinical endpoints were investigated using a random forest model following 70%-30% training-test split. For the DL analysis, both 2D and 3D model designs were executed using two different deep neural networks as transfer learning problems: in 2D-based design, 8x8cm2 axial fields-of-view(FOVs) centered within the GTV were adopted for VGG-16 training; in 3D-based design, 8x8x8 cm3 FOVs centered within the GTV were adopted for U-Net’s encoder path training. In both designs, data augmentation (rotation, translation, flip, noise) was included to overcome potential training convergence problems due to the imbalanced dataset, and the same 70%-30% training-test split was used. The performances of the 3 models (Radiomics, 2D-DL, 3D-DL) were tested to predict outcomes including local failure, non-local failure, and disease-free survival. Sensitivity/specificity/accuracy/ROC results were obtained from their 20 trained versions. Results: The radiomics models showed limited performances in all three outcome prediction tasks. The 2D-DL design showed significant improvement compared to the radiomics results in predicting local failure (ROC AUC = 0.546±0.056). The 3D-DL design achieved the best performance for all three outcomes (local failure ROC AUC = 0.768 ± 0.051, non-local failure ROC AUC = 0.683±0.027, disease-free ROC AUC = 0.694±0.042) with statistically significant improvements from radiomics/2D-DL results. Conclusions: 3D-DL execution outperformed the 2D-DL in predicting clinical outcomes after surgery for early-stage NSCLC. By contrast, classic radiomics approach did not achieve satisfactory results.

AbstractPurpose: To develop a deep ensemble learning model with a radiomics spatial encoding execution for improved glioma segmentation accuracy using multi-parametric MRI (mp-MRI). Materials/Methods: This radiomics-incorporated deep ensemble learning model was developed using 369 glioma patients with a 4-modality mp-MRI protocol: T1, contrast-enhanced T1 (T1-Ce), T2, and FLAIR. In each modality volume, a 3D sliding kernel was implemented across the brain to capture image heterogeneity: fifty-six radiomic features were extracted within the kernel, resulting in a 4th order tensor. Each radiomic feature can then be encoded as a 3D image volume, namely a radiomic feature map (RFM). For each patient, all RFMs extracted from all 4 modalities were processed by the Principal Component Analysis (PCA) for dimension reduction, and the first 4 principal components (PCs) were selected. Next, four deep neural networks following the U-net’s architecture were trained for the segmenting of a region-of-interest (ROI): each network utilizes the mp-MRI and 1 of the 4 PCs as a 5-channel input for 2D execution. Last, the 4 softmax probability results given by the U-net ensemble were superimposed and binarized by Otsu’s method as the segmentation result. Three deep ensemble models were trained to segment enhancing tumor (ET), tumor core (TC), and whole tumor (WT), respectively. Segmentation results given by the proposed ensemble were compared to the mp-MRI-only U-net results. Results: All 3 radiomics-incorporated deep learning ensemble models were successfully implemented: Compared to mp-MRI-only U-net results, the dice coefficients of ET (0.777→0.817), TC (0.742→0.757), and WT (0.823→0.854) demonstrated improvements. Accuracy, sensitivity, and specificity results demonstrated the same patterns. Conclusion: The adopted radiomics spatial encoding execution enriches the image heterogeneity information that leads to the successful demonstration of the proposed neural network ensemble design, which offers a new tool for mp-MRI-based medical image segmentation.

Background: Radiomics is a process of converting medical images into high-dimensional quantitative features and the subsequent mining these features for providing decision support. It is conducted as a potential noninvasive, low-cost, and patient-specific routine clinical tool. Building a predictive model which is reliable, efficient, and accurate is a vital part for the success of radiomics. Machine learning method is a powerful tool to achieve this. Feature extraction strongly affects the performance. Delta-feature is one way of feature extraction methods to reflect the spatial variation in tumor phenotype, hence it could provide better treatment-specific assessment.

Purpose: To compare the performance of using pre-treatment features and delta-features for assessing the brain radiosurgery treatment response, and to investigate the performance of different combinations of machine learning methods for feature selection and for feature classification.

Materials and Methods: A cohort of 12 patients with brain treated by radiosurgery was included in this research. The pre-treatment, one-week post-treatment, and two-month post-treatment T1 and T2 FLAIR MR images were acquired. 61 radiomic features were extracted from the gross tumor volume (GTV) of each image. The delta-features from pre-treatment to two post-treatment time points were calculated. With leave-one-out sampling, pre-treatment features and the two sets of delta-features were separately input into a univariate Cox regression model and a machine learning model (L1-regularized logistic regression [L1-LR], random forest [RF] or neural network [NN]) for feature selection. Then a machine learning method (L1-LR, L2-regularized logistic regression [L2-LR], RF, NN, kernel support vector machine [Kernel-SVM], linear-SVM, or naïve bayes [NB]) was used to build a classification model to predict overall survival. The performance of each model combination and feature type was estimated by the area under receiver operating characteristic (ROC) curve (AUC).

Results: The AUC of one-week delta-features was significantly higher than that of pre-treatment features (p-values < 0.0001) and two-month delta-features (p-value= 0.000). The model combinations of L1-LR for feature selection and RF for classification as well as RF for feature selection and NB for classification based on one-week delta-features presented the highest AUC values (both AUC=0.944).

Conclusions: This work potentially implied that the delta-features could be better in predicting treatment response than pre-treatment features, and the time point of computing the delta-features was a vital factor in assessment performance. Analyzing delta-features using a suitable machine learning approach is potentially a powerful tool for assessing treatment response.

Purpose: To develop a framework to quantify the impact of different CT imaging parameters on the variations of radiomic features and to investigate the effectiveness of different image processing methods on the reduction of radiomic feature variations.

Method: A publicly available CT image dataset (Credence Cartridge Radiomics Phantom CT scans) acquired on a phantom consisting of different texture materials was used in this study. 251 scans were divided into 5 groups. In each group, only one of the following imaging parameters changed: slice thickness, pixel size, convolutional kernel, exposure (mAs), and scanner model. 92 radiomic features from intensity, intensity histogram, GLCM, GLRLM, and GLSZM groups were extracted from the same region of interest (ROI) in each scan using an in-house application. The coefficient of variation (CV) was used to measure the variation of radiomic features due to each imaging parameter. Three preprocessing methods, resampling, gray level rebinning, and image filtering were tested for their effectiveness in reducing feature variations.

Result: The proposed workflow identified individual features and groups of features with high variation and showed responses of each feature to different image preprocessing methods. The convolutional kernel of scanners caused the largest variations in calculated features in general, while exposure had low influence on features in all categories. GLSZM features showed higher sensitivity to pixel size and slice thickness due to the dependence of number of voxels in a gray level zone to the voxel size. Image preprocessing did not improve feature robustness in most cases.

Conclusion: This study demonstrated the ability to reveal the relationships between radiomic feature variations, imaging parameters, and image correction methods. The proposed workflow can be used to study the feature robustness prior to the application of any radiomic features in multi-institutional studies.

Purpose: To develop a novel voxel-based radiomics extraction technique, and to investigate the potential association between spatially-encoded radiomics features of the lungs and pulmonary function.

Methods: We developed a voxel-based radiomics feature extraction platform to generate radiomics filtered images. Specifically, for each voxel in the image, 62 radiomics features were calculated in a rotationally-invariant 3D neighbourhood to capture spatially-encoded information. In general, such an approach results in an image tensor object, i.e., each voxel in the original image is represented by a 62-dimensional radiomics feature vector. Two digital phantoms are then designed to validate the technique's ability to quantify regional image information. To test the technique as a potential pulmonary biomarker, we generated radiomics filtered images for 25 lung CT image and are subsequently evaluated against corresponding Galligas PET images, as the ground truth for pulmonary function, using voxel-wise Spearman correlation (r). The Canonical Correlation Analysis (CCA)-based feature fusion method is also implemented to enhance such a correlation. Finally, the Spearman distributions were compared with 37 individual CT ventilation image (CTVI) algorithms to assess the overall performance relative to conventional CT-based techniques.

Results: Several radiomics filtered images were identified to be correlated with Galligas PET lung imaging. The most robust association was found to be the Run Length Encoding feature, Run-Length Non-uniformity (0.21

Conclusions: This preliminary study indicates that spatially-encoded lung texture and lung density are potentially associated with pulmonary function as measured via Galligas PET ventilation images. Collectively, low density, heterogeneous coarse lung texture was often associated with lower Galligas radiotracer amounts.

Purpose: To characterize the effect of different discretization techniques (methods and values) on radiomic features extracted from positron emission tomography (PET) images of patients undergoing definitive radiotherapy for oropharyngeal cancer (OPC) and determine if there are optimal binning techniques associated with the computed texture and histogram measurements.Methods: 71 patients were enrolled in a prospective clinical trial to receive definitive radiotherapy (70Gy) for OPC. PET/CT images were acquired both prior to treatment and two weeks into treatment (i.e., after 20 Gy). All patients were scanned on the same PET/CT imaging system. The gross tumor volume at the primary tumor site was manually segmented on CT and transferred to PET, from which 74 quantitative radiomic features were extracted as potential imaging biomarkers. The sensitivity of feature extraction to common discretization techniques (fixed bin number vs. fixed bin size) was systematically evaluated by measuring radiomic feature values at monotonically increasing bin numbers (32, 64, 128, 256) and bin sizes (0.1, 0.5, 1.0, 5.0). Disparities in radiomics data parameterized by these different discretization settings were quantified based on t-tests of individual features and cross-correlation of matrix-level feature spaces. A discretization invariance score (DIS) was defined as an aggregation of each unique probability of rejecting the null hypothesis that any two discretization techniques produce the same feature value. To evaluate the generalization of these characteristics during treatment, DIS values were compared between pre- and intra-treatment imaging. Results: Only 50% of radiomic features were robust (DIS > 0.7) to changes in bin number, compared to 66% of features when varying bin size. Regardless of discretization technique, grey level variance (DIS=0.0) and high grey level size emphasis (DIS=0.21) were the most sensitive to binning perturbations, while skewness (DIS=1.0) and kurtosis (DIS=1.0) were nearly invariant. The cross-correlation between discretization-specific feature spaces was maximized for fixed bin number and minimized for fixed bin size. Ranked DIS measurements were comparable between pre-treatment and intra-treatment imaging, implying that feature sensitivity is invariant to changes in the absolute feature value over time. Conclusion: The impact of discretization is largely feature-dependent. Individual features demonstrated a non-linear response to systematic changes in discretization parameters, which was captured by our DIS metric. DIS values can be used to optimize downstream radiomic biomarkers, where the prognostic value of individual features may depend on feature-specific discretization.

Purpose: There is an urgent need for enhanced prognostic tools and insights into the biology andbiomarkers of chemoradiation resistance. The purpose of this research is to identify prognostic radiomic features on 18F-FDG micro-PET/CT images for the response to chemoradiation in mouse models of head and neck squamous cell carcinoma (HNSCC). Methods: Two orthotopic murine human papillomavirus (HPV)-negative (MOC1, MOC2) models and one HPV-positive HNSCC model (MLM1) were utilized. Oral cavity tumors were induced by injecting HNSCC cells into the buccal mucosa of C57BL/6J mice. Bidirectional caliper tumor measurements were conducted thrice weekly with chemoradiation initiated once tumors exceeded 50mm3: cisplatin (5 mg/kg, intraperitoneal) and image-guided radiation therapy (8 Gy) on days 0 and 7. On day 14, 18F-FDG PET/CT imaging was performed. Mice were euthanized when they reached humane endpoint (tumor >12mm, any dimension). Tumors were segmented on PET/CT, and volume, SUVmean, and SUVmax were extracted. Liver regions of interest were segmented for normalization of tumor SUVmax to liver SUVmean. Treatment response was evaluated using tumor size on day 10 relative to day 0. Tumor growth and survival were compared across models (Kruskal-Wallis with Tukey’s post hoc test) and based on image feature parameters (Mann-Whitney U test). The associations between survival, SUVmax, and tumor volumes were analyzed by Cox proportional hazards model and Kaplan-Meier curves with log-rank test. Results: 121 mice underwent treatment and imaging. Univariate analysis showed that median tumor volume and SUVmax were significantly associated with survival and treatment response (p<0.05). The Cox model indicated a significant difference in survival probability based on risk score values derived from the model's coefficients estimating their relative risk of time-to-event (p<0.0001). Conclusion: These results suggest that image features on 18F-PET/CT can provide prognostic insight into treatment response and survival in preclinical HNSCC models, providing a platform for further studies to improve understanding of the biological underpinnings of radiomic expression associated with chemoradiation resistance.

Artificial intelligence (AI), including classic machine learning (ML) and deep learning (DL), has recently made an impact on advanced medical image analysis. Classic ML learns the data representation by manual image feature engineering, namely radiomics, based on experts' domain knowledge. DL directly learns the image feature through hierarchical data modeling directly from the input data. Both classic ML and DL models have emerged as promising AI tools for medical image analysis. Despite promising academic research in which algorithms are beginning to outperform humans, clinical radiography analysis still has limited AI involvement. One issue of current AI development (for both classic ML and DL) is the lack of model explainability, i.e., the extent to which the internal mechanics of an AI model can be explained in human terms from a clinical perspective. The unexplainable issues include, but are not limited to, model confidence ('Can we trust the results with some clues?'), data utilization ('Do we need this as a part of the model?'), and model generalization ('How do I know if it works?'). Without such model explainability, AI models remain a black box in implementation, which leads to a lack of accountability and confidence in clinic application. We hypothesize that the current medical domain knowledge, both in theory and in practice, can be incorporated into AI designs to provide explainability. Therefore, the objective of this dissertation is to explore potential techniques to enhance AI model explainability. Specifically, three novel AI models were developed: • The first model aimed to explore a radiomic filtering model to quantify and visualize radiomic features associated with pulmonary ventilation from lung computed tomography (CT). In this model, lung volume was segmented on 46 CT images, and a 3D sliding window kernel was implemented across the lung volume to capture the spatial-encoded image information. Fifty-three radiomic features were extracted within the kernel, resulting in a 4th-order tensor object. As such, each voxel coordinate of the original lung was represented as a 53-dimensional feature vector, such that radiomic features could be viewed as feature maps within the lungs. To test the technique as a potential pulmonary ventilation biomarker, the radiomic feature maps were compared to paired functional images (Galligas-positron emission tomography, PET or DTPA-single photon emission computed tomography, SPECT) based on Spearman correlation (?) analysis. From the results, the radiomic feature map Gray Level Run Length Matrix (GLRLM)-based Run-Length Non-Uniformity and Gray Level Co-occurrence Matrix (GLCOM)-based Sum Average are found to be highly correlated with functional imaging. The achieved ? (median [range]) for the two features are 0.46 [0.05, 0.67] and 0.45 [0.21, 0.65] across 46 patients and 2 functional imaging modalities, respectively. The results provide evidence that local regions of sparsely encoded heterogeneous lung parenchyma on CT are associated with diminished radiotracer uptake and measured lung ventilation defects on PET/SPECT imaging. Collectively, these findings demonstrate the potential of radiomic filtering to provide a visual explanation of lung CT radiomic features associated with lung ventilation. The developed technique may serve as a complementary tool to the current lung quantification techniques and provide hypothesis-generating data for future studies. • The second model aimed to explore a neural ordinary differential equation (ODE)-based segmentation model to observe deep neural network (DNN) behavior in multi-parametric magnetic resonance imaging (MRI)-based glioma segmentation. In this model, by hypothesizing that deep feature extraction can be modeled as a spatiotemporally continuous process, we implemented a novel DL model, neural ODE, in which deep feature extraction was governed by an ODE parameterized by a neural network. The dynamics of 1) MR images after interactions with the DNN and 2) segmentation formation can thus be visualized after solving the ODE. An accumulative contribution curve (ACC) was designed to quantitatively evaluate each MR image’s utilization by the DNN toward the final segmentation results. The proposed neural ODE model was demonstrated using 369 glioma patients with a 4-modality multi-parametric MRI protocol: T1, contrast-enhanced T1 (T1-Ce), T2, and fluid-attenuated inversion recovery (FLAIR). Three neural ODE models were trained to segment enhancing tumor (ET), tumor core (TC), and whole tumor (WT), respectively. The key MR modalities with significant utilization by DNNs were identified based on ACC analysis. Segmentation results by DNNs using only the key MR modalities were compared to the ones using all 4 MR modalities in terms of Dice coefficient, accuracy, sensitivity, and specificity. From the results, all neural ODE models successfully illustrated image dynamics as expected. ACC analysis identified T1-Ce as the only key modality in ET and TC segmentations, while both FLAIR and T2 were key modalities in WT segmentation. Compared to the U-Net results using all 4 MR modalities, the Dice coefficient of ET (0.784→0.775), TC (0.760→0.758), and WT (0.841→0.837). Collectively, the neural ODE model offers a new tool for optimizing the DL model inputs with enhanced explainability in data utilization. The presented methodology can be generalized to other medical image-related DL applications. • The third model aimed to explore a multi-feature-combined (MFC) model to quantify the role of radiomic features, DL image features, and their combination in predicting local failure from pre-treatment CT images of early-stage non-small cell lung cancer (NSCLC) patients after either lung surgery or stereotactic body radiation therapy (SBRT). The MFC model comprised three key steps. (1) Extraction of 92 handcrafted radiomic features from the gross tumor volume (GTV) segmented on pre-treatment CT images. (2) Extraction of 512 deep features from pre-trained DL U-Net encoder structure. Specifically, the 512 latent activation values from the last fully connected layers were studied. (3) The extracted 92 handcrafted radiomic features, 512 deep features, along with 4 patient demographic information (i.e., gender, age, tumor volume, and Charlson comorbidity index), were concatenated as a multi-dimensional input to three classifiers: logistic regression (LR), supporting vector machine (SVM), and random forest (RF) to predict the local failure. Two NSCLC patient cohorts from our institution were investigated: (1) the surgery cohort includes 83 patients who underwent segmentectomy or wedge resection (with 7 local failures), and (2) the SBRT cohort includes 84 patients who received lung SBRT (with 9 local failures). The MFC model was developed and evaluated independently for both patient cohorts. For each cohort, the MFC model was also compared against (1) the R model: LR/SVM/RF prediction models using only radiomic features, (2) the PI model: LR/SVM/RF prediction models using only patient demographic information, and (3) the DL model: DL design that directly predicts the local failure based on the U-Net encoder. All models were tested based on two validation methods: leave-one-out cross-validation (LOOCV) and 100-fold Monte Carlo cross-validation (MCCV) with a 70%-30% train-test ratio. ROC with AUC analysis was adopted as the main evaluator to measure the prediction performance. The student’s t-test was performed to identify the statistically significant differences when applicable. In LOOCV, the AUC range of the proposed MFC model (for three classifiers) was 0.811-0.956 for the surgery patient cohort and 0.913-0.981 for the SBRT cohort, which was higher than the other studied models: the AUC range was 0.356-0.480 (surgery) and 0.295-0.347 (SBRT) for the PI models, 0.388-0.655 (surgery) and 0.648-0.747 (SBRT) for the R models, and 0.816 (surgery) and 0.842 (SBRT) for the DL models. Similar results can be observed in the 100-fold MCCV: the MFC model again showed the highest AUC results (surgery: 0.831-0.841, SBRT: 0.860-0.947), which were significantly higher than the PI models (surgery: 0.464-0.564, SBRT: 0.457-0.519), R models (surgery: 0.546-0.653, SBRT: 0.559-0.667), and DL models (surgery: 0.690, SBRT: 0.773). Collectively, the developed MFC model improves the ability to predict the occurrence of local failure for both surgery and SBRT patient cohorts with enhanced explainability in the role of different feature sources. It may hold the potential to assist clinicians to optimize treatment procedures in the future. In summary, the three developed models provide substantial contributions to enhance the explainability of current classic ML and DL models. The concepts and techniques developed in this dissertation, as well as understandings and inspirations from the key results, provide valuable knowledge for the future development of AI techniques toward wide clinical trust and acceptance.

Purpose: To investigate the feasibility of integrate radiomics and deep learning in computer-aided medical imaging analysis Methods: Two different approaches were investigated to integrate radiomics and deep leaning on two independent tasks respectively. In the first approach, a 2D sliding kernel was implemented to map the impulse response of radiomic features throughout the entire chest X-ray image; thus, each feature is rendered as a 2D map in the same dimension as the X-ray image. Based on each of the three investigated deep neural network architectures, including VGG-16, VGG-19, and DenseNet-121, a pilot model was trained using X-ray images only. Subsequently, 2 radiomic feature maps (RFMs) were selected based on cross-correlation analysis in reference to the pilot model saliency map results. The radiomics-boosted model was then trained based on the same deep neural network architecture using X-ray images plus the selected RFMs as input. The proposed radiomics-boosted design was developed using 812 chest X-ray images with 262/288/262 COVID-19/Non-COVID-19 pneumonia/healthy cases, and 649/163 cases were assigned as training-validation/independent test sets. For each model, 50 runs were trained with random assignments of training/validation cases following the 7:1 ratio in the training-validation set. Sensitivity, specificity, accuracy, and ROC curves together with Area-Under-the-Curve (AUC) from all three deep neural network architectures were evaluated. In the second approach, a cohort of 235 GBM patients with complete surgical resection was divided into short-term/long-term survival groups with 1-yr survival time threshold. Each patient received a pre-surgery multi-parametric MRI exam with 4 scans: T1, contrast-enhanced T1 (T1ce), T2, and FLAIR. Three tumor subregions were segmented by neuroradiologists, and the whole dataset was divided into training, validation, and test groups following a 7:1:2 ratio. The developed model comprises three data source branches: in the 1st radiomics branch, 456 radiomics features (RF) were calculated from the three tumor subregions of each patient’s MR images; in the 2nd deep learning branch, an encoding neural network architecture was trained for survival group prediction using each single MR modality, and high-dimensional parameters from the last two network layers were extracted as deep features (DF). The extracted radiomics features and deep features were processed by a feature selection procedure to reduce the dimension size of each feature space. In the 3rd branch, patient-specific clinical features (PSCF), including patient age and three tumor subregions volumes, were collected from the dataset. Finally, data sources from all three branches were fused as an integrated input for a supporting vector machine (SVM) execution for survival group prediction. Different strategies of model design were investigated in comparison studies, including 1) 2D/3D-based image analysis, 2) different radiomics feature space dimension reduction methods, and 3) different data source combinations in SVM input design. Results: In the first approach, all three investigated deep neural network architectures demonstrated improved sensitivity, specificity, accuracy, and ROC AUC results in COVID-19 and healthy individual classifications. VGG-16 showed the largest improvement in COVID-19 classification ROC (AUC from 0.963 to 0.993), and DenseNet-121 showed the largest improvement in healthy individual classification ROC (AUC from 0.962 to 0.989). The reduced variations suggested improved robustness of the model to data partition. For the challenging Non-COVID-19 pneumonia classification task, radiomics-boosted implementation of VGG-16 (AUC from 0.918 to 0.969) and VGG-19 (AUC from 0.964 to 0.970) improved ROC results, while DenseNet-121 showed a slight yet insignificant ROC performance reduction (AUC from 0.963 to 0.949). The achieved highest accuracy of COVID-19/Non-COVID-19 pneumonia/healthy individual classifications were 0.973 (VGG-19)/0.936 (VGG-19)/ 0.933 (VGG-16), respectively. In the second approach, the model achieved 0.638 prediction accuracy in the test set when using patient-specific clinical features only, which was higher than the results using radiomics features/deep features as sole input of SVM in both 2D and 3D based analysis. The inclusion of radiomics features or deep features with patient-specific clinical features improved accuracy results in 3D analysis. The most accurate models in 2D/3D analysis reached the highest accuracy of 0.745 with different combinations of dissimilarity-selected radiomics features, deep features, and patient-specific clinical features, and the corresponding ROC area-under-curve (AUC) results were 0.69 (2D) and 0.71 (3D), respectively.

Conclusions: The integration of radiomic analysis in deep learning model design improved the performance and robustness computer-aided diagnosis and outcome predication, which holds great potential for clinical applications and provides a radiomics perspective for deep learning interpretation.

As modern diagnostic imaging systems become increasingly more quantitative, new techniques and scientific disciplines are emerging as powerful avenues to personalized medicine. Leading this paradigm shift is the field of radiomics, which attempts to identify computational biomarkers hidden within high-throughput imaging data. Radiomic biomarkers may be able to non-invasively detect the underlying phenotype of an image, leading to new insights and innovation. Such insights may include correlations between radiomic features and pathological information, treatment response, functional characteristics, etc. Searching for meaningful structure within these quantitative datasets is therefore fundamental to contemporary imaging science.

However, imaging data is being created at an alarming rate, and the ability to understand hyper-dimensional relationships between radiomic features is often non-trivial. This is a major challenge for radiomic applications to clinical medicine. There is an urgent need to investigate novel technologies to manage this challenge, so that radiomics can be effectively and efficiently used to solve complex clinical problems.

Major contributions of this dissertation research include: (1) the development of a novel data clustering algorithm called Langevin annealing; (2) the development of a translational research environment to use this clustering algorithm for oncological imaging characterization; and (3) applications of the developed technique for clinical diagnosis and treatment evaluation.

Cluster analysis – i.e., the grouping of similar data objects together based on their intrinsic properties – is a common approach to understanding otherwise non-trivial data. Although data clustering is a hallmark of many fields, it is generally an ill-defined practice. Notable limitations and challenges include: (1) defining the appropriate number of clusters, (2) poor optimization near local minima, and (3) black-box approaches that often make interpretation difficult.

To overcome some of these challenges, data clustering may benefit from physics intuition. Langevin Annealing models radiomics data as a dynamical system in equilibrium with a heat bath. The method is briefly summarized as follows. (1) A radial basis function is used to construct a density distribution, , from the radiomics data. (2) A potential, , is then constructed such that is the ground-state solution to the time-independent Schrödinger equation. (3) Using , Langevin dynamics are formulated at sub-critical temperature to avoid ergodicity, and the radiomic feature vectors are propagated as the system evolves.

The time dynamics of individual radiomic feature vectors lead to different metastable states, which are interpreted as clusters. Clustering is achieved when subsets of the data aggregate near minima of . While the radiomic feature vectors are pushed towards potential minima by the potential gradient, , Brownian motion allows them to effectively tunnel through local potential barriers and escape saddle points into functional locations of the potential surface otherwise forbidden. Nearly degenerate local minima can merge, allowing hyper-dimensional radiomics data to be explored at high resolution, while still maintaining a reasonably narrow impulse response.

Since radiomics is still a rather immature field, there is currently a lack of commercially-available software. Therefore, a radiomic feature extraction platform was developed to facilitate this dissertation research. The extraction code – which is the primary focus of Chapter 2 – is the means to converting unstructured data (i.e., images) into structured data (i.e., features). It therefore serves as a translational research environment that provides the necessary input to subsequent radiomic analyses.

Imaging features derived from dynamic environments – such as the lungs – are highly susceptible to variability and motion artifacts. Before implementing major analyses and new techniques, Chapter 3 investigates the spatial-temporal variability of radiomic features. This problem is approached based on computational experiments using both (a) a simulated dynamic digital phantom and (b) real patient CT data. Key findings demonstrate that radiomic features are sensitive to spatial-temporal changes, which may influence the quality of feature analyses. In general, radiomic feature-sensitivity is shown to be broad and inherently feature-specific.

The theory and development of Langevin annealing is covered in Chapter 4, where a complete theory and mathematical derivation is formulated. Several illustrating examples and computational simulations are used to demonstrate the clustering technique. Chapter 5 provides a comprehensive validation of Langevin annealing using a common benchmark dataset. Accurate ergodic sampling is achieved, clustering performance is evaluated, hyper-parameters are characterized, and the approach is compared to several well-known clustering algorithms.

While this dissertation has broader application to many aspects of medical imaging, a majority of the analysis is conducted on patients with non-small cell lung cancer (NSCLC). In particular, two classes of CT-based radiomic biomarkers are considered throughout this work: (1) radiomic biomarkers derived from lungs, and (2) radiomic biomarkers derived from tumors. These radiomic biomarkers are characterized in Chapter 6 and Chapter 7, where the emphasis of the dissertation shifts from highly theoretical work to a more application-driven focus.

Radiomic lung biomarkers are investigated in Chapter 6, where several associations are identified linking the quantitative imaging data to pulmonary function. In general, patients with larger lungs of homogeneous, low attenuating pulmonary tissue are shown to have worse pulmonary function. Radiomic tumor biomarkers are investigated in Chapter 7, where several associations are identified linking the quantitative imaging data to treatment response. In general, relatively dense tumors with a homogenous coarse texture are shown to be linked with higher rates of local cancer recurrence following stereotactic body radiation therapy.

Background: Quantitative medical imaging has been increasingly utilized in modern medicine. The field of radiomics is an emerging subset of quantitative medical imaging. Radiomics can identify a large number of quantitative features as the biomarkers from the radiography images. Subsequent mining and analysis of these features may potentially uncover subclinical tumor characteristics of disease in a non-invasive manner, and provide clinical decision support. Biomarkers on gross tumor regions have been studied for many years for different clinical endpoints. Correlations between the radiomic features and treatment outcomes, disease diagnosis, pathological information, and organ functions can be studied for clinical decision support. Peritumor region is the region around the gross tumor, where may also be used to extract radiomics biomarkers. Currently, limited studies are available in this area. Machine learning is a powerful tool to correlate radiomics features with various clinical endpoints and provide a predictive model to better define tumor characteristics.

Purpose: To investigate the association between CT radiomic data from peritumoral regions and non-local treatment failure recurrence of early stage non-small cell lung cancers (NSCLC) following lung stereotactic body radiation therapy (SBRT), to identify secondary quantitative parameters from the dose-driven peritumoral CT radiomic data, to compare the model performance of using radiomic data with and without secondary quantitative parameters for assessing the non-local treatment failure following lung SBRT.

Materials and Methods: Sixty-three patients who received SBRT for early-stage NSCLC were retrospectively identified by an IRB approved clinical trial with treatment outcome provided. Treatment failure was defined as both local cancer recurrence and non-local cancer recurrence following SBRT. Gross Tumor Volumes (GTVs) were segmented on the pre-treatment free-breathing CT images by radiation oncologists. Two types of peritumoral volumes were defined on pre-treatment free-breathing CT: uniform ring and dose-driven. The uniform ring peritumoral volumes were generated by expanding GTVs at radial distances of 3mm, 6mm, 9mm and 12mm. The dose-driven peritumoral volumes were generated by converting ten isodose volumes (100%, 98%, 95%, 90%, 85%, 80%, 70%, 50%, 30%, and 10%) into structures in the treatment planning system. All peritumoral volumes were then modified using a Boolean process to exclude the GTV and non-lung tissue. Sixty radiomic features (4 Intensity, 21 GLCOM, 11 GLRLM, 13 GLSZM, 5 NGLDM, and 6 Shape) were extracted from GTVs and peritumoral volumes using an in-house radiomics calculation platform as biomarkers for cancer recurrence. A univariate feature selection was used to eliminate highly correlated features. Two multivariate machine learning based feature selection algorithms were followed to find the important features and to avoid overfitting. Machine learning algorithms (Random Forest, LASSO Logistic Regression, Ridge Logistic Regression, Linear Support Vector Machine and Kernel Support Vector Machine) were used to build classification models to predict non-local treatment failure. 63 patients were randomly separated into the training group (75%) and the test group (25%). Model generalization was cross validated using a stratified 10-fold with 50 iterations. The learning curve and grid search were used to find the best hyperparameters to enhance the model performance and avoid the overfitting in the random forest model. The performance of each feature selection and classifier combination was based on the area under the receiver operating characteristic curve (ROC). Pair t-tests with 100 iteration permutation tests were used to evaluate the significant difference between different AUCs.

Results: In uniform expansion ring peritumoral region, from the results, 3mm ring was more predictive of non-local treatment failure than all peritumoral radiomic data and gross tumor radiomic data. In dose-driven peritumoral region, from the results, 80% isodose region was more predictive of non-local treatment failure than all peritumoral radiomic data and gross tumor radiomic data. The model combinations of tree-based feature selection and Random Forest classifier presented the highest AUC values. No significant difference in predictive performance was found before and after adding secondary features in the dataset (p-value > 0.05). However, secondary features may be useful in other studies given that obvious patterns were seen in the clustering heatmap.

Conclusions: This study has demonstrated strong prognostic value of peritumoral radiomics for non-local treatment failure in patients with stage I NSCLC. The presented peritumoral radiomics was shown to have better predictive performance compared to the gross tumor radiomics.

Keywords: Radiomics, peritumoral region, machine learning, stereotactic body radiation therapy, non-small cell lung cancer

Lung cancer in accounts for 13% of all new cancer diagnoses and is the leading cause of cancer mortality in the United States (Bogart, 2017, Howlader, 2020). Non-small cell lung cancer (NSCLC) in particular, accounts for 80-85% of lung cancer diagnoses and is estimated to cause more than 130,000 deaths this year (American Cancer Society, 2021). Currently, the standard of care for early-stage NSCLC is surgery, with stereotactic body radiation therapy (SBRT) is becoming more accepted as the primary treatment option for patients who are medically inoperable. It remains controversial as to which method is optimal for marginal surgical patients, but it has been shown that SBRT and sublobar resection provide similar local tumor control rates and clinical outcomes in stage I NSCLC (Ackerson, 2018).The goal of this research work was to develop pre-treatment radiomic models for surgical NSCLC patients to predict cancer recurrence. This was done by investigating two specific aims. The first was to (1.) build radiomic models based on pre-treatment CT images from surgical patients and evaluate their performance in predicting cancer recurrence and the second was to (2.) build radiomic models based on pre-treatment CT images from surgical patients and evaluate their performance in predicting cancer recurrence. Radiomic features were extracted from the contoured GTV’s from pre-treatment CT scans of surgical and SBRT patients. To investigate the first aim, multivariate models were trained and tested on only surgical patients to find associations between the extracted features and each clinical outcome. To investigate the second aim, these models were first trained on surgical data and tested on SBRT data to investigate the generalizability of each model across treatment modalities. Next, models were trained and tested on a pooled dataset to investigate potential associations of radiomic features with cancer recurrence independent of treatment. Models were evaluated by creating ROC curves and calculating the area under these curves (AUC’s). Models trained and tested on surgical patients showed a stronger association between radiomic features and non-local failure (maximum AUC of 0.82 ± 0.04) and a poor association with local failure (maximum AUC of 0.57 ± 0.04). This may suggest that radiomic features have limited value in predicting local recurrence since the GTV which is used in calculating these features is no longer in the body post-treatment. Despite this, it is difficult to draw strong conclusions based on the variability in the image parameters of surgical patients, such as slice thickness and x-ray tube current, which have been shown to affect feature values (Midya, 2018, Kim, 2019). This is supported by the degraded performance in these models when SBRT data was introduced, further increasing image variability.

Purpose: To develop a novel radiomic quantification framework to quantify the impact of image resampling and discretization on radiomic texture characterization performance.

Methods: The study employed 251 CT scans of a Credence Cartridge phantom (consisting of 10 texture materials) with different image acquisition parameters. Each material was segmented using a pre-defined cylindrical mask. Different image pre-processing workflows including 5 resampling methods (no resampling, trilinear, and nearest resampling to both 1mm³ and 5mm³) and 8 discretization methods (fixed bin size of 25,50,75,100 and fixed bin counts of 8,16,32,64) were randomly applied. 75 radiomic texture features (including 24GLCM-based, 16GLRLM-based, 16GLSZM-based, 14GLDM-based, and 5NGTDM-based) were extracted from each material to characterize its textural attributes. Three machine learning models including logistic regression (LR), random forest (RF), and supporting vector machine (SVM) were developed to identify 10 materials based on the extracted features, and grid search was adopted to optimize the model hyperparameters. The model performance was evaluated on 10-class macro-AUC with 5-fold cross-validation.

Results: Three models successfully classified 10 materials with macro-AUC=0.9941±0.0081, 0.9979±0.0040, and 0.9957±0.0067 for LR, RF, and SVM, respectively. Across 8 different discretization methods, an increasing trend in performance can be observed when the original CT was discretized to a larger gray level range: performance improved by 0.0038 with bin sizes decreasing from 100-25, and by 0.0074 with bin counts increasing from 8 to 64. Among 5 resampling methods, resampling CT to an isotropic voxel spacing showed an improved prediction performance (0.9942±0.0075/0.9944±0.0073 for trilinear/nearest resampling to 1mm³ and 5mm³, respectively) over no interpolation (0.9862±0.0228), with minimal performance discrepancies observed among two different interpolation algorithms. In addition, no statistically significant differences were observed across five folds.

Conclusion: The proposed framework successfully quantified the dependence of radiomics texture characterization on image resampling and discretization.

Background: Radiomics describes the study of converting medical images into high-dimensional quantitative features and following analysis for further decision making and genomics focuses on the understanding genomes of individual organisms and characterizations of different genomes. Radiogenomics is a new emerging method that combines both radiomics and genomics together in clinical studies as well as researches the relation of genetic characteristics and radiomic features. It has the potential as a tool for medical treatment assessment in the future. In this study, we used machine learning methods to build two models for treatment assessment: 1) the output is p53 mutation, and the inputs are radiomic features; 2) the output is patient overall survival, and the inputs are radiomic features and p53 mutation. The modelling process was divided into feature selection and classification. Machine learning is a popular area of artificial intelligence that can make machines “learn by itself”. Machine learning algorithm learns from datasets called “training data”, and generates a prediction model from its learning process. The prediction model can then be used to make predictions and decisions from other datasets.

Purpose: 1) To investigate the correlation between p53 mutation and radiomic features in lung cancer, and to detect p53 mutation from radiomic features using different machine learning methods, 2) To investigate the correlation between genomic (p53 mutation), radiomic, radiogenomic features and overall patient survival in lung cancer using machine learning methods.

Material and Methods: The study used 24 patients with advanced lung cancers who had received radiotherapy and chemotherapy. CT was used as medical imaging modality in radiomics study. A radiomics study was then performed which involved three parts: Pre-treatment (Pre-Tx) Radiomics, Post-treatment (Post-Tx) Radiomics, and Delta Radiomics. The pre-Tx radiomic features were calculated from treatment planning CT images, the post-Tx radiomic features were calculated from the follow-up CT images after the radiotherapy, and the delta radiomic features were calculated as the change of radiomic features between cancer treatment. 19 of 24 patients had both pre-Tx and post-Tx CT images. Totally 61 representative radiomic features were extracted from CT images, including Intensity features, Grey Level Co-occurrence Matrix features, Grey Level Run Length Matrix Features, Grey Level Size Zone Matrix features, Neighborhood Grey Level Difference Matrix features, and Morphological features. Feature selection was implemented to avoid feature redundancy. Spearman Correlation analysis and Lasso regression were used for feature selection for p53 mutation detection. Cox regression and lasso regression were used for feature selection for patient survival prediction. Then, several common machine learning based classification methods were used for modelling of p53 mutation detection and patient survival prediction, including linear discriminative analysis, quadratic discriminative analysis, Naïve Bayes, Linear Support Vector Machine, Kernel Support Vector Machine, Bootstrap Aggregating (Bagging), Logistic Regression, and Lasso generalized linear regression. Radiomic models were used for p53 mutation detection in tumor. Radiogenomic models based on combined radiomic features and p53 mutation were used for patient overall survival prediction. To avoid bias, the leave-one-out cross validation method was used for both feature selection and classification. Receiver Operator Characteristic (ROC) Curves were used as an evaluation method for the model, and Area Under Curve (AUC) values were compared for different classification methods.

Results: For p53 mutation detection, the highest AUC of pre-Tx radiomics (24 patients), pre-Tx radiomics (19 patients), and post-Tx radiomics (19 patients) was 0.6993, 0.5606, and 0.6591. For patient survival prediction, the highest AUC of pre-Tx radiomics (24 patients), pre-Tx radiomics (19 patients), post-Tx radiomics (19 patients), and delta radiomics (19 patients) was 0.7045, 0.7125, 0.6063, and 0.8000, and the highest AUC of pre-Tx radiogenomics (24 patients), pre-Tx radiogenomics (19 patients), post-Tx radiogenomics (19 patients), and delta radiogenomics (19 patients) was 0.7500, 0.7375, 0.5857, and 0.9143.

Conclusion: From limited dataset, it might be feasible to detect p53 mutation by both pre-Tx and post-Tx radiomics. Lasso and LSVM has shown the best performance in classification.

For predicting the overall patient survival, different features were selected. This may be related to the limited data available for the study. It may also be related to the different characteristics of pre-Tx, post-Tx and delta radiomics. Intensity and texture features showed high frequency being selected for pre-Tx and delta features, and morphological features showed high frequency for post-Tx radiomics. However, we also found that the combination of delta radiomics and p53 mutation showed a better patient survival prediction than pre-Tx, post-Tx, delta radiomics and p53 mutation alone. The reason might be related to the difference of tumor reaction to radiation due to p53 mutation. KSVM and Bagging showed highest performance compared with other classification methods.

Keyword: Radiogenomics, radiomics, delta radiomics, genomics, p53, lung cancer, radiotherapy.

This study aims to (1) investigate methodology for harmonization of radiomics features between planning CT and on-board CBCT and establish a workflow to harmonize images taken from different scanning protocols and over the course of radiotherapy treatments using normalization, and (2) examine feature variability of longitudinal cone-beam CT radiomics for 3 different fractionation schemes and a time-point during treatment indicative of early treatment response.

All CBCT images acquired over the course of lung SBRT for each patient were registered with corresponding planning CT. A volume-of-interest (VOI) in a homogeneous soft-tissue region that would not change over the course of radiotherapy was selected on the planning CT. The VOI was applied to all CBCT images of the same patient taken at different days. The first CBCT was normalized to the planning CT using the ratio of their respective mean VOI pixel values. Subsequent CBCT images were normalized using the ratio of that CBCT’s mean VOI pixel value to the first CBCT’s mean VOI pixel value. Forty-three features characterizing image intensity and morphology in fine and coarse textures were extracted from the planning CT, all original CBCT images, and all normalized CBCT images. T-test on extracted features from CBCT images with and without normalization indicates the effect of normalization on the distribution of various features. Mutual information between the planning CT and the first CBCT with and without normalization was calculated to assess the effectiveness of normalization on harmonizing radiomics features.

Of 72 NSCLC patients treated with lung SBRT, 18 received 15-18 Gy / fraction for 3 fractions; 36 received 12-12.5 Gy / fraction for 4 fractions; 18 received 8-10 Gy / fraction for 5 fractions. We studied 7 sets of CBCT images from the same treatment fraction as a ‘test-retest’ baseline to study the additional daily CBCT images. Fifty-five gray level intensity and textural features were extracted from the CBCT images. Test-retest images were used to determine the smallest detectable change (C=1.96*SD) indicating significant variation with a 95% confidence level. Here, the significance of feature variation depended on the choice of a minimum number of patients for which a feature changed more than ’C’. Analysis of which features change at which moment during treatment with different fractionation schemes was used to investigate a time-point indicative of early tumor response.

T-test on planning CT and CBCT images of the 72 patients indicated that normalization with a soft tissue VOI reduced the number of features with significant variation (p<0.05) by 55%. Following lung SBRT, 30 features changed significantly for at least 10% of all patients. For patients treated with 3 fractions, 49 features changed at Fraction 2, and 49 at Fraction 3; there was 100% overlap between features at both fractions. For patients treated with 4 fractions, 45, 45, and 48 features changed at Fraction 2-4 respectively; there was 92% overlap between features at Fraction 2 and the remaining fractions. For patients treated with 5 fractions, 12, 18, 14, and 25 features changed at Fraction 2-5; there was 36%, 48%, and 48% overlap between features at Fraction 2-4 and the remaining fractions respectively.

Normalization can potentially harmonize radiomics features on both planning CT and on-board CBCT. Feature variability depends on the selection of normalization VOI and extraction VOI. Significant changes in gray level radiomic features were observed over the course of lung SBRT. Different fractionation schemes corresponded to different patterns of feature variation. Higher fractional dose corresponded to a larger number of variable features and high overlap of variable features at an earlier time-point.

The purpose of this work was to develop a new radiomics paradigm for sparse, time-series imaging data, where features are extracted from a spatial-temporal manifold modeling the time evolution between images, and to assess the prognostic value on patients with oropharyngeal cancer (OPC).To accomplish this, we developed an algorithm to mathematically describe the relationship between two images acquired at time t=0 and t>0. These images serve as boundary conditions of a partial differential equation describing the transition from one image to the other. To solve this equation, we propagate the position and momentum of each voxel according to Fokker-Planck dynamics (i.e., a technique common in statistical mechanics). This transformation is driven by an underlying potential force uniquely determined by the equilibrium image. The solution generates a spatial-temporal manifold (3 spatial dimensions + time) from which we define dynamic radiomic features. First, our approach was numerically verified by stochastically sampling dynamic Gaussian processes of monotonically decreasing noise. The transformation from high to low noise was compared between our Fokker-Planck estimation and simulated ground-truth. To demonstrate feasibility and clinical impact, we applied our approach to 18F-FDG-PET images to estimate early metabolic response of patients (n=57) undergoing definitive (chemo)radiation for OPC. Images were acquired pre-treatment and two-weeks intra-treatment (after 20 Gy). Dynamic radiomic features capturing changes in texture and morphology were then extracted. Patients were partitioned into two groups based on similar dynamic radiomic feature expression via k-means clustering and compared by Kaplan-Meier analyses with log-rank tests (p<0.05). These results were compared to conventional delta radiomics to test the added value of our approach. Numerical results confirmed our technique can recover image noise characteristics given sparse input data as boundary conditions. Our technique was able to model tumor shrinkage and metabolic response. While no delta radiomics features proved prognostic, Kaplan-Meier analyses identified nine significant dynamic radiomic features. The most significant feature was Gray-Level-Size-Zone-Matrix gray-level variance (p=0.011), which demonstrated prognostic improvement over its corresponding delta radiomic feature (p=0.722). We developed, verified, and demonstrated the prognostic value of a novel, physics-based radiomics approach over conventional delta radiomics via data assimilation of quantitative imaging and differential equations.

The purpose of this dissertation was to improve CT-based radiomics characterization by assessing and accounting for its systematic and stochastic variability due to variations in the imaging method. The anatomically informed methodologies developed in this dissertation enable radiomics studies to retrospectively correct for the effects CT imaging protocols and prospectively inform CT protocol choices. This project was conducted in three parts of 1) assessing of the bias and variability of morphologic radiomics features across a wide range of CT imaging protocols and segmentation algorithms, 2) assessing the applicability, sensitivity, and usefulness of applying bias correction factors retrospectively to patient data acquired with heterogenous CT imaging protocols, and 3) developing analytical techniques to reduce the variability of radiomics features by prospectively optimizing the CT imaging protocols.

In part 1 (chapters 2-4), the measurability of bias and variability of morphologic radiomcis features was assessed. In chapter 2, a theoretical framework was developed to guide the process of analyzing and utilizing quantitative features, including radiomics, derived from CT images. The framework outlined the key qualities necessary for successful quantification including biological and clinical relevance, objectivity, robustness, and implementability. In chapter 3, a method was developed to use anatomically informed lung lesion models to assess the bias and variability of morphology radiomics features as a function of CT imaging protocols and segmentation algorithms. The results showed that bias and variability of radiomics features are dependent on a complicated interplay of anatomical, imaging protocol, and segmentation effects. In chapter 4, the bias and variability of radiomics due to segmentation algorithms was explored in-depth for three segmentation algorithms across a range of image noise magnitudes. The segmentation algorithms were assessed by comparing their performance to an ideal radiomics estimator for a range of image quality characteristics. The results showed that the optimal segmentation algorithm was function of the specific noise magnitude and the radiomics features of interest.

In part 2 (chapter 5), an analysis was carried out using a Non-Small Cell Lung Cancer patient dataset to assess the applicability, sensitivity, and usefulness of correcting radiomics features for imaging protocol effects. The applicability was assessed by calculating bias correction factors from one set of anatomically informed lesion models and applying the correction factors to another set of anatomically informed lesion models. The sensitivity was assessed by applying idealized bias correction factors to the patient dataset with increasing bias correction magnitudes to determine the sensitivity of predictive models to the magnitude of the bias correction factors. Finally, the usefulness was assessed by applying the anatomically informed protocol-specific bias correction factors to the patient dataset and quantifying the change in the performance predictive model. The results showed that the bias correction factors are applicable when the bias correction factors are derived from and applied to lesion models with similar anatomical characteristics. The feature-specific sensitivity of prediction to bias correction factors was found to be as low as 1-5% and was typically in the range of 20-50%. The bias correction factors were applied to a patient population and were found to result in a small statistically significant improvement in the performance.

In part 3 (chapter 6), a method was developed and implemented to assess the minimum detectable difference of morphologic radiomics features as a function of protocol and anatomical characteristics. The analysis of the data was carried out to allow for evaluating and informing the recommendations of the Quantitative Imaging Biomarkers Alliance (QIBA) for lung nodule volumetry. The results showed that the minimum detectable difference for QIBA compliant protocols was a lower median value than the minimum detectable difference among all possible CT protocols. The techniques developed in this analysis can be used to prospectively optimize CT imaging protocols for improved quantitative characterization of radiomics features.

In conclusion, this dissertation developed methods to assess and account for the variability of radiomics features across CT imaging protocols and segmentation algorithms using anatomically informed lesion models.