NRG Oncology Assessment of Artificial Intelligence Deep Learning-Based Auto-segmentation for Radiation Therapy: Current Developments, Clinical Considerations, and Future Directions.

Deep learning neural networks (DLNN) in Artificial intelligence (AI) have been extensively explored for automatic segmentation in radiotherapy (RT). In contrast to traditional model-based methods, data-driven AI-based models for auto-segmentation have shown high accuracy in early studies in research settings and controlled environment (single institution). Vendor-provided commercial AI models are made available as part of the integrated treatment planning system (TPS) or as a stand-alone tool that provides streamlined workflow interacting with the main TPS. These commercial tools have drawn clinics' attention thanks to their significant benefit in reducing the workload from manual contouring and shortening the duration of treatment planning. However, challenges occur when applying these commercial AI-based segmentation models to diverse clinical scenarios, particularly in uncontrolled environments. Contouring nomenclature and guideline standardization has been the main task undertaken by the NRG Oncology. AI auto-segmentation holds the potential clinical trial participants to reduce interobserver variations, nomenclature non-compliance, and contouring guideline deviations. Meanwhile, trial reviewers could use AI tools to verify contour accuracy and compliance of those submitted datasets. In recognizing the growing clinical utilization and potential of these commercial AI auto-segmentation tools, NRG Oncology has formed a working group to evaluate the clinical utilization and potential of commercial AI auto-segmentation tools. The group will assess in-house and commercially available AI models, evaluation metrics, clinical challenges, and limitations, as well as future developments in addressing these challenges. General recommendations are made in terms of the implementation of these commercial AI models, as well as precautions in recognizing the challenges and limitations.

Prediction of Radiation Treatment Response for Locally Advanced Rectal Cancer via a Longitudinal Trend Analysis Framework on Cone-Beam CT.

Locally advanced rectal cancer (LARC) presents a significant challenge in terms of treatment management, particularly with regards to identifying patients who are likely to respond to radiation therapy (RT) at an individualized level. Patients respond to the same radiation treatment course differently due to inter- and intra-patient variability in radiosensitivity. In-room volumetric cone-beam computed tomography (CBCT) is widely used to ensure proper alignment, but also allows us to assess tumor response during the treatment course. In this work, we proposed a longitudinal radiomic trend (LRT) framework for accurate and robust treatment response assessment using daily CBCT scans for early detection of patient response. The LRT framework consists of four modules: (1) Automated registration and evaluation of CBCT scans to planning CT; (2) Feature extraction and normalization; (3) Longitudinal trending analyses; and (4) Feature reduction and model creation. The effectiveness of the framework was validated via leave-one-out cross-validation (LOOCV), using a total of 840 CBCT scans for a retrospective cohort of LARC patients. The trending model demonstrates significant differences between the responder vs. non-responder groups with an Area Under the Curve (AUC) of 0.98, which allows for systematic monitoring and early prediction of patient response during the RT treatment course for potential adaptive management.

Feature selection based on unsupervised clustering evaluation for predicting neoadjuvant chemoradiation response for patients with locally advanced rectal cancer.

Accurate response prediction allows for personalized cancer treatment of locally advanced rectal cancer (LARC) with neoadjuvant chemoradiation. In this work, we designed a convolutional neural network (CNN) feature extractor with switchable 3D and 2D convolutional kernels to extract deep learning features for response prediction. Compared with radiomics features, convolutional kernels may adaptively extract local or global image features from multi-modal MR sequences without the need of feature predefinition. We then developed an unsupervised clustering based evaluation method to improve the feature selection operation in the feature space formed by the combination of CNN features and radiomics features. While normal process of feature selection generally includes the operations of classifier training and classification execution, the process needs to be repeated many times after new feature combinations were found to evaluate the model performance, which incurs a significant time cost. To address this issue, we proposed a cost effective process to use a constructed unsupervised clustering analysis indicator to replace the classifier training process by indirectly evaluating the quality of new found feature combinations in feature selection process. We evaluated the proposed method using 43 LARC patients underwent neoadjuvant chemoradiation. Our prediction model achieved accuracy, area-under-curve (AUC), sensitivity and specificity of 0.852, 0.871, 0.868, and 0.735 respectively. Compared with traditional radiomics methods, the prediction models (AUC = 0.846) based on deep learning-based feature sets are significantly better than traditional radiomics methods (AUC = 0.714). The experiments also showed following findings: (1) the features with higher predictive power are mainly from high-order abstract features extracted by CNN on ADC images and T2 images; (2) both ADC_Radiomics and ADC_CNN features are more advantageous for predicting treatment responses than the radiomics and CNN features extracted from T2 images; (3) 3D CNN features are more effective than 2D CNN features in the treatment response prediction. The proposed unsupervised clustering indicator is feasible with low computational cost, which facilitates the discovery of valuable solutions by highlighting the correlation and complementarity between different types of features.

MRI-Guided Radiation Therapy for Prostate Cancer: The Next Frontier in Ultrahypofractionation.

Technological advances in MRI-guided radiation therapy (MRIgRT) have improved real-time visualization of the prostate and its surrounding structures over CT-guided radiation therapy. Seminal studies have demonstrated safe dose escalation achieved through ultrahypofractionation with MRIgRT due to planning target volume (PTV) margin reduction and treatment gating. On-table adaptation with MRI-based technologies can also incorporate real-time changes in target shape and volume and can reduce high doses of radiation to sensitive surrounding structures that may move into the treatment field. Ongoing clinical trials seek to refine ultrahypofractionated radiotherapy treatments for prostate cancer using MRIgRT. Though these studies have the potential to demonstrate improved biochemical control and reduced side effects, limitations concerning patient treatment times and operational workflows may preclude wide adoption of this technology outside of centers of excellence. In this review, we discuss the advantages and limitations of MRIgRT for prostate cancer, as well as clinical trials testing the efficacy and toxicity of ultrafractionation in patients with localized or post-prostatectomy recurrent prostate cancer.

Predicting Overall Survival for Patients with Malignant Mesothelioma Following Radiotherapy via Interpretable Machine Learning.

PURPOSE/OBJECTIVES: Malignant pleural mesothelioma (MPM) is a rare but aggressive cancer arising from the cells of the thoracic pleura with a poor prognosis. We aimed to develop a model, via interpretable machine learning (ML) methods, predicting overall survival for MPM following radiotherapy based on dosimetric metrics as well as patient characteristics. MATERIALS/METHODS: Sixty MPM (37 right, 23 left) patients treated on a Tomotherapy unit between 2013 and 2018 were retrospectively analyzed. All patients received 45 Gy (25 fractions). The multivariable Cox regression (Cox PH) model and Survival Support Vector Machine (sSVM) were applied to build predictive models of overall survival (OS) based on clinical, dosimetric, and combined variables. RESULTS: Significant differences in dosimetric endpoints for critical structures, i.e., the lung, heart, liver, kidney, and stomach, were observed according to target laterality. The OS was found to be insignificantly different (p = 0.18) between MPM patients who tested left- and right-sided, with 1-year OS of 77.3% and 75.0%, respectively. With Cox PH regression, considering dosimetric variables for right-sided patients alone, an increase in PTV_Min, Total_Lung_PTV_Mean, Contra_Lung_Volume, Contra_Lung_V20, Esophagus_Mean, and Heart_Volume had a greater hazard to all-cause death, while an increase in Total_Lung_PTV_V20, Contra_Lung_V5, and Esophagus_Max had a lower hazard to all-cause death. Considering clinical variables alone, males and increases in N stage had greater hazard to all-cause death; considering both clinical and dosimetric variables, increases in N stage, PTV_Mean, PTV_Min, and esophagus_Mean had greater hazard to all-cause death, while increases in T stage and Heart_V30 had lower hazard to all-cause-death. In terms of C-index, the Cox PH model and sSVM performed similarly and fairly well when considering clinical and dosimetric variables independently or jointly. CONCLUSIONS: Clinical and dosimetric variables may predict the overall survival of mesothelioma patients, which could guide personalized treatment planning towards a better treatment response. The identified predictors and their impact on survival offered additional value for translational application in clinical practice.

A subregion-based survival prediction framework for GBM via multi-sequence MRI space optimization and clustering-based feature bundling and construction.

Aiming at accurate survival prediction of Glioblastoma (GBM) patients following radiation therapy, we developed a subregion-based survival prediction framework via a novel feature construction method on multi-sequence MRIs. The proposed method consists of two main steps: (1) a feature space optimization algorithm to determine the most appropriate matching relation derived between multi-sequence MRIs and tumor subregions, for using multimodal image data more reasonable; (2) a clustering-based feature bundling and construction algorithm to compress the high-dimensional extracted radiomic features and construct a smaller but effective set of features, for accurate prediction model construction. For each tumor subregion, a total of 680 radiomic features were extracted from one MRI sequence using Pyradiomics. Additional 71 geometric features and clinical information were collected resulting in an extreme high-dimensional feature space of 8231 to train and evaluate the survival prediction at 1 year, and the more challenging overall survival prediction. The framework was developed based on 98 GBM patients from the BraTS 2020 dataset under five-fold cross-validation, and tested on an external cohort of 19 GBM patients randomly selected from the same dataset. Finally, we identified the best matching relationship between each subregion and its corresponding MRI sequence, a subset of 235 features (out of 8231 features) were generated by the proposed feature bundling and construction framework. The subregion-based survival prediction framework achieved AUCs of 0.998 and 0.983 on the training and independent test cohort respectively for 1 year survival prediction, compared to AUCs of 0.940 and 0.923 for survival prediction using the 8231 initial extracted features for training and validation cohorts respectively. Finally, we further constructed an effective stacking structure ensemble regressor to predict the overall survival with the C-index of 0.872. The proposed subregion-based survival prediction framework allow us to better stratified patients towards personalized treatment of GBM.

A multi-objective based radiomics feature selection method for response prediction following radiotherapy.

Objective.Radiomics contains a large amount of mineable information extracted from medical images, which has important significance in treatment response prediction for personalized treatment. Radiomics analyses generally involve high dimensions and redundant features, feature selection is essential for construction of prediction models.Approach.We proposed a novel multi-objective based radiomics feature selection method (MRMOPSO), where the number of features, sensitivity, and specificity are jointly considered as optimization objectives in feature selection. The MRMOPSO innovated in the following three aspects: (1) Fisher score to initialize the population to speed up the convergence; (2) Min-redundancy particle generation operations to reduce the redundancy between radiomics features, a truncation strategy was introduced to further reduce the number of features effectively; (3) Particle selection operations guided by elitism strategies to improve local search ability of the algorithm. We evaluated the effectiveness of the MRMOPSO by using a multi-institution oropharyngeal cancer dataset from The Cancer Imaging Archive. 357 patients were used for model training and cross validation, an additional 64 patients were used for evaluation.Main results.The area under the curve (AUC) of our method achieved AUCs of 0.82 and 0.84 for cross validation and independent dataset, respectively. Compared with classical feature selection methods, the AUC of MRMOPSO is significantly higher than the Lasso (AUC = 0.74,p-value = 0.02), minimal-redundancy-maximal-relevance criterion (mRMR) (AUC = 0.73,p-value = 0.05), F-score (AUC = 0.48,p-value < 0.01), and mutual information (AUC = 0.69,p-value < 0.01) methods. Compared to single-objective methods, the AUC of MRMOPSO is 12% higher than those of the genetic algorithm (GA) (AUC = 0.68,p-value = 0.02) and particle swarm optimization algorithm (AUC = 0.72,p-value = 0.05) methods. Compared to other multi-objective feature selection methods, the AUC of MRMOPSO is 14% higher than those of multiple objective particle swarm optimization (MOPSO) (AUC = 0.68,p-value = 0.02) and nondominated sorting genetic algorithm II (NSGA2) (AUC = 0.70,p-value = 0.03).Significance.We proposed a multi-objective based radiomics feature selection method. Compared to conventional feature reduction algorithms, the proposed algorithm effectively reduced feature dimension, and achieved superior performance, with improved sensitivity and specificity, for response prediction in radiotherapy.

Quality and Safety Considerations in Image Guided Radiation Therapy: An ASTRO Safety White Paper Update.

PURPOSE: This updated report on image guided radiation therapy (IGRT) is part of a series of consensus-based white papers previously published by the American Society for Radiation Oncology addressing patient safety. Since the first white papers were published, IGRT technology and procedures have progressed significantly such that these procedures are now more commonly used. The use of IGRT has now extended beyond high-precision treatments, such as stereotactic radiosurgery and stereotactic body radiation therapy, and into routine clinical practice for many treatment techniques and anatomic sites. Therefore, quality and patient safety considerations for these techniques remain an important area of focus. METHODS AND MATERIALS: The American Society for Radiation Oncology convened an interdisciplinary task force to assess the original IGRT white paper and update content where appropriate. Recommendations were created using a consensus-building methodology, and task force members indicated their level of agreement based on a 5-point Likert scale from "strongly agree" to "strongly disagree." A prespecified threshold of ≥75% of raters who selected "strongly agree" or "agree" indicated consensus. SUMMARY: This IGRT white paper builds on the previous version and uses other guidance documents to primarily focus on processes related to quality and safety. IGRT requires an interdisciplinary team-based approach, staffed by appropriately trained specialists, as well as significant personnel resources, specialized technology, and implementation time. A thorough feasibility analysis of resources is required to achieve the clinical and technical goals and should be discussed with all personnel before undertaking new imaging techniques. A comprehensive quality-assurance program must be developed, using established guidance, to ensure IGRT is performed in a safe and effective manner. As IGRT technologies continue to improve or emerge, existing practice guidelines should be reviewed or updated regularly according to the latest American Association of Physicists in Medicine Task Group reports or guidelines. Patient safety in the application of IGRT is everyone's responsibility, and professional organizations, regulators, vendors, and end-users must demonstrate a clear commitment to working together to ensure the highest levels of safety.

AAPM Task Group Report 306: Quality control and assurance for tomotherapy: An update to Task Group Report 148.

Since the publication of AAPM Task Group (TG) 148 on quality assurance (QA) for helical tomotherapy, there have been many new developments on the tomotherapy platform involving treatment delivery, on-board imaging options, motion management, and treatment planning systems (TPSs). In response to a need for guidance on quality control (QC) and QA for these technologies, the AAPM Therapy Physics Committee commissioned TG 306 to review these changes and make recommendations related to these technology updates. The specific objectives of this TG were (1) to update, as needed, recommendations on tolerance limits, frequencies and QC/QA testing methodology in TG 148, (2) address the commissioning and necessary QA checks, as a supplement to Medical Physics Practice Guidelines (MPPG) with respect to tomotherapy TPS and (3) to provide risk-based recommendations on the new technology implemented clinically and treatment delivery workflow. Detailed recommendations on QA tests and their tolerance levels are provided for dynamic jaws, binary multileaf collimators, and Synchrony motion management. A subset of TPS commissioning and QA checks in MPPG 5.a. applicable to tomotherapy are recommended. In addition, failure mode and effects analysis has been conducted among TG members to obtain multi-institutional analysis on tomotherapy-related failure modes and their effect ranking.

Radiation therapy practice changes in the COVID-19 pandemic era: A pilot study in California.

PURPOSE: This study aims to investigate practice changes among Southern and Northern California's radiation oncology centers during the COVID-19 pandemic. METHODS: On the online survey platform SurveyMonkey, we designed 10 survey questions to measure changes in various aspects of medical physics practice. The questions covered patient load and travel rules; scopes to work from home; new protocols to reduce corona virus disease-2019 (COVID-19) infection risk; availability of telemedicine; and changes in fractionation schedules and/or type of treatment plans. We emailed the survey to radiation oncology centers throughout Northern and Southern California, requesting one completed survey per center. All responses were anonymized, and data were analyzed using both qualitative and quantitative research methods. RESULTS: At the end of a 4-month collection period (July 2, 2021 to October 11, 2021), we received a total of 61 responses throughout Southern and Northern California. On average, 4111 patients were treated per day across the 61 centers. New COVID-19-related department and hospital policies, along with hybrid workflow changes, infectious control policies, and changes in patient load have been reported. Results also showed changes in treatment methods during the pandemic, such as increased use of telemedicine, hypofractionation for palliative, breast cancer, and prostate cancer cases; and simultaneous boosts, compared to sequential boosts. CONCLUSION: Our California radiation oncology center population study shows changes in various aspects of radiation oncology practices during the COVID-19 pandemic. This study serves as a pilot study to identify possible correlations and new strategies that allow radiation oncology centers to continue providing quality patient care while ensuring the safety of both staff and patients.

Bridging Feature Gaps to Improve Multi-Organ Segmentation on Abdominal Magnetic Resonance Image.

Accurate segmentation of abdominal organs on MRI is crucial for computer-aided surgery and computer-aided diagnosis. Most state-of-the-art methods for MRI segmentation employ an encoder-decoder structure, with skip connections concatenating shallow features from the encoder and deep features from the decoder. In this work, we noticed that simply concatenating shallow and deep features was insufficient for segmentation due to the feature gap between shallow features and deep features. To mitigate this problem, we quantified the feature gap from spatial and semantic aspects and proposed a spatial loss and a semantic loss to bridge the feature gap. The spatial loss enhanced spatial details in deep features, and the semantic loss introduced semantic information into shallow features. The proposed method successfully aggregated the complementary information between shallow and deep features by formulating and bridging the feature gap. Experiments on two abdominal MRI datasets demonstrated the effectiveness of the proposed method, which improved the segmentation performance over a baseline with nearly zero additional parameters. Particularly, the proposed method has advantages for segmenting organs with blurred boundaries or in a small scale, achieving superior performance than state-of-the-art methods.

A hybrid feature selection-based brain tumor detection and segmentation in multiparametric magnetic resonance imaging.

PURPOSE: To develop a novel method based on feature selection, combining convolutional neural network (CNN) and ensemble learning (EL), to achieve high accuracy and efficiency of glioma detection and segmentation using multiparametric MRIs. METHODS: We proposed an evolutionary feature selection-based hybrid approach for glioma detection and segmentation on 4 MR sequences (T2-FLAIR, T1, T1Gd, and T2). First, we trained a lightweight CNN to detect glioma and mask the suspected region to process large batch of MRI images. Second, we employed a differential evolution algorithm to search a feature space, which composed of 416-dimensions radiomics features extracted from four sequences of MRIs and 128-dimensions high-order features extracted by the CNN, to generate an optimal feature combination for pixel classification. Finally, we trained an EL classifier using the optimal feature combination to segment whole tumor (WT) and its subregions including non-enhancing tumor (NET), peritumoral edema (ED), and enhancing tumor (ET) in the suspected region. Experiments were carried out on 300 glioma patients from the BraTS2019 dataset using fivefold cross-validation, and the model was independently validated using the rest 35 patients from the same database. RESULTS: The approach achieved a detection accuracy of 98.8% using four MRIs. The Dice coefficients (and standard deviations) were 0.852 ± 0.057, 0.844 ± 0.046, and 0.799 ± 0.053 for segmentation of WT (NET+ET+ED), tumor core (NET+ET), and ET, respectively. The sensitivities and specificities were 0.873 ± 0.074, 0.863 ± 0.072, and 0.852 ± 0.082; the specificities were 0.994 ± 0.005, 0.994 ± 0.005, and 0.995 ± 0.004 for the WT, tumor core, and ET, respectively. The performances and calculation times were compared with the state-of-the-art approaches, our approach yielded a better overall performance with average processing time of 139.5 s per set of four sequence MRIs. CONCLUSIONS: We demonstrated a robust and computational cost-effective hybrid segmentation approach for glioma and its subregions on multi-sequence MR images. The proposed approach can be used for automated target delineation for glioma patients.

A hybrid feature selection-based approach for brain tumor detection and automatic segmentation on multiparametric magnetic resonance images.

PURPOSE: To develop a novel method based on feature selection, combining convolutional neural network (CNN) and ensemble learning (EL), to achieve high accuracy and efficiency of glioma detection and segmentation using multiparametric MRIs. METHODS: We proposed an evolutionary feature selection-based hybrid approach for glioma detection and segmentation on 4 MR sequences (T2-FLAIR, T1, T1Gd, and T2). First, we trained a lightweight CNN to detect glioma and mask the suspected region to process large batch of MRI images. Second, we employed a differential evolution algorithm to search a feature space, which composed of 416-dimension radiomic features extracted from four sequences of MRIs and 128-dimension high-order features extracted by the CNN, to generate an optimal feature combination for pixel classification. Finally, we trained an EL classifier using the optimal feature combination to segment whole tumor (WT) and its subregions including nonenhancing tumor (NET), peritumoral edema (ED), and enhancing tumor (ET) in the suspected region. Experiments were carried out on 300 glioma patients from the BraTS2019 dataset using fivefold cross validation, the model was independently validated using the rest 35 patients from the same database. RESULTS: The approach achieved a detection accuracy of 98.8% using four MRIs. The Dice coefficients (and standard deviations) were 0.852 ± 0.057, 0.844 ± 0.046, and 0.799 ± 0.053 for segmentation of WT (NET+ET+ED), tumor core (NET+ET), and ET, respectively. The sensitivities and specificities were 0.873 ± 0.074, 0.863 ± 0.072, and 0.852 ± 0.082; the specificities were 0.994 ± 0.005, 0.994 ± 0.005, and 0.995 ± 0.004 for the WT, tumor core, and ET, respectively. The performances and calculation times were compared with the state-of-the-art approaches, our approach yielded a better overall performance with average processing time of 139.5 s per set of four sequence MRIs. CONCLUSIONS: We demonstrated a robust and computational cost-effective hybrid segmentation approach for glioma and its subregions on multi-sequence MR images. The proposed approach can be used for automated target delineation for glioma patients.

Three-dimensional multipath DenseNet for improving automatic segmentation of glioblastoma on pre-operative multimodal MR images.

PURPOSE: Convolutional neural networks have achieved excellent results in automatic medical image segmentation. In this study, we proposed a novel three-dimensional (3D) multipath DenseNet for generating the accurate glioblastoma (GBM) tumor contour from four multimodal pre-operative MR images. We hypothesized that the multipath architecture could achieve more accurate segmentation than a singlepath architecture. METHODS: Two hundred and fifty-eight GBM patients were included in this study. Each patient had four MR images (T1-weighted, contrast-enhanced T1-weighted, T2-weighted, and FLAIR) and the manually segmented tumor contour. We built a 3D multipath DenseNet that could be trained to achieve an end-to-end mapping from four MR images to the corresponding GBM tumor contour. A 3D singlepath DenseNet was also built for comparison. Both DenseNets were based on the encoder-decoder architecture. All four images were concatenated and fed into a single encoder path in the singlepath DenseNet, while each input image had its own encoder path in the multipath DenseNet. The patient cohort was randomly split into a training set of 180 patients, a validation set of 39 patients, and a testing set of 39 patients. Model performance was evaluated using the Dice similarity coefficient (DSC), average surface distance (ASD), and 95% Hausdorff distance (HD95% ). Wilcoxon signed-rank tests were conducted to assess statistical significances. RESULTS: The singlepath DenseNet achieved the DSC of 0.911 ± 0.060, ASD of 1.3 ± 0.7 mm, and HD95% of 5.2 ± 7.1 mm, while the multipath DenseNet achieved the DSC of 0.922 ± 0.041, ASD of 1.1 ± 0.5 mm, and HD95% of 3.9 ± 3.3 mm. The P-values of all Wilcoxon signed-rank tests were less than 0.05. CONCLUSIONS: Both DenseNets generated GBM tumor contours in good agreement with the manually segmented contours from multimodal MR images. The multipath DenseNet achieved more accurate tumor segmentation than the singlepath DenseNet. Here presented the 3D multipath DenseNet that demonstrated an improved accuracy over comparable algorithms in the clinical task of GBM tumor segmentation.

Clinical Development and Evaluation of Megavoltage Topogram for Fast Patient Alignment on Helical Tomotherapy.

PURPOSE: To develop and evaluate a fast patient localization tool using megavoltage (MV)-topogram on helical tomotherapy. METHODS AND MATERIALS: Eighty-one MV-topogram pairs for 18 pelvis patients undergoing radiation were acquired weekly under an institutional review board-approved clinical trial. The MV-topogram imaging protocol requires 2 orthogonal acquisitions at static gantry angles of 0 degrees and 90 degrees for a programed scan length. A MATLAB based in-house software was developed to reconstruct the MV-topograms offline. Reference images (digitally reconstructed topograms, digitally reconstructed topograms) were generated using the planning computed tomography and tomotherapy geometry. The MV-topogram based alignment was determined by registering the MV-topograms to the digitally reconstructed topogram using bony landmark on commercial MIM software. The daily shifts in 3 translational directions determined from MV-topograms were compared with the megavoltage computed tomography (MVCT) based patient shifts. Linear-regression and two one-sided tests equivalence tests were performed to investigate the relation and equivalence between the 2 techniques. Seventy-eight MV-topogram pairs for 19 head and neck patients were included to validate the finding. RESULTS: The magnitudes of alignment differences of (MVCT - MV-topogram) (and standard deviations) were -0.3 ± 2.1, -0.8 ± 2.4, and 1.6 ± 1.7 mm for pelvis and 0.6 ± 1.2, 0.8 ± 4.2, 1.6 ± 2.6 mm for head and neck; the linear-regression coefficients between 2 imaging techniques were 1.18, 1.10, 0.94, and 0.86, 0.63, 0.38 in the lateral, longitudinal, vertical directions for pelvis and head and neck, respectively. The acquisition time for a pair of MV-topograms was up to 12.7 times less than MVCT scans (coarse scan mode) while covering longer longitudinal length. CONCLUSIONS: MV-topograms showed equivalent clinical performance to the standard MVCT with significantly less acquisition time for pelvis and H&N patients. The MV-topogram can be used as an alternative or complimentary tool for bony landmark-based patient alignment on tomotherapy.

Development and Validation of a Comprehensive Multivariate Dosimetric Model for Predicting Late Genitourinary Toxicity Following Prostate Cancer Stereotactic Body Radiotherapy.

Purpose: Dosimetric predictors of toxicity after Stereotactic Body Radiation Therapy (SBRT) are not well-established. We sought to develop a multivariate model that predicts Common Terminology Criteria for Adverse Events (CTCAE) late grade 2 or greater genitourinary (GU) toxicity by interrogating the entire dose-volume histogram (DVH) from a large cohort of prostate cancer patients treated with SBRT on prospective trials. Methods: Three hundred and thirty-nine patients with late CTCAE toxicity data treated with prostate SBRT were identified and analyzed. All patients received 40 Gy in five fractions, every other day, using volumetric modulated arc therapy. For each patient, we examined 910 candidate dosimetric features including maximum dose, volumes of each organ [CTV, organs at risk (OARs)], V100%, and other granular volumetric/dosimetric indices at varying volumetric/dosimetric values from the entire DVH as well as ADT use to model and predict toxicity from SBRT. Training and validation subsets were generated with 90 and 10% of the patients in our cohort, respectively. Predictive accuracy was assessed by calculating the area under the receiver operating curve (AROC). Univariate analysis with student t-test was first performed on each candidate DVH feature. We subsequently performed advanced machine-learning multivariate analyses including classification and regression tree (CART), random forest, boosted tree, and multilayer neural network. Results: Median follow-up time was 32.3 months (range 3-98.9 months). Late grade ≥2 GU toxicity occurred in 20.1% of patients in our series. No single dosimetric parameter had an AROC for predicting late grade ≥2 GU toxicity on univariate analysis that exceeded 0.599. Optimized CART modestly improved prediction accuracy, with an AROC of 0.601, whereas other machine learning approaches did not improve upon univariate analyses. Conclusions: CART-based machine learning multivariate analyses drawing from 910 dosimetric features and ADT use modestly improves upon clinical prediction of late GU toxicity alone, yielding an AROC of 0.601. Biologic predictors may enhance predictive models for identifying patients at risk for late toxicity after SBRT.

Survival prediction for oral tongue cancer patients via probabilistic genetic algorithm optimized neural network models.

OBJECTIVES: High throughput pre-treatment imaging features may predict radiation treatment outcome and guide individualized treatment in radiotherapy (RT). Given relatively small patient sample (as compared with high dimensional imaging features), identifying potential prognostic imaging biomarkers is typically challenging. We aimed to develop robust machine learning methods for patient survival prediction using pre-treatment quantitative CT image features for a subgroup of head-and-neck cancer patients. METHODS: Three neural network models, including back propagation (BP), Genetic Algorithm-Back Propagation (GA-BP), and Probabilistic Genetic Algorithm-Back Propagation (PGA-BP) neural networks were trained to simulate association between patient survival and radiomics data in radiotherapy. To evaluate the models, a subgroup of 59 head-and-neck patients with primary cancers in oral tongue area were utilized. Quantitative image features were extracted from planning CT images, a novel t-Distributed Stochastic Neighbor Embedding (t-SNE) method was used to remove irrelevant and redundant image features before fed into the network models. 80% patients were used to train the models, and remaining 20% were used for evaluation. RESULTS: Of the three supervised machine-learning methods studied, PGA-BP yielded the best predictive performance. The reported actual patient survival interval of 30.5 ± 21.3 months, the predicted survival times were 47.3 ± 38.8, 38.5 ± 13.5 and 29.9 ± 15.3 months using the traditional PCA. Combining with the novel t-SNE dimensionality reduction algorithm, the predicted survival intervals are 35.8 ± 15.2, 32.3 ± 13.1 and 31.6 ± 15.8 months for the BP, GA-BP and PGA-BP neural network models, respectively. CONCLUSION: The work demonstrated that the proposed probabilistic genetic algorithm optimized neural network models, integrating with the t-SNE dimensionality reduction algorithm, achieved accurate prediction of patient survival. ADVANCES IN KNOWLEDGE: The proposed PGA-BP neural network, integrating with an advanced dimensionality reduction algorithm (t-SNE), improved patient survival prediction accuracy using pre-treatment quantitative CT image features of head-and-neck cancer patients.

Multi-to-binary network (MTBNet) for automated multi-organ segmentation on multi-sequence abdominal MRI images.

Fully convolutional neural network (FCN) has achieved great success in semantic segmentation. However, the performance of the FCN is generally compromised for multi-object segmentation. Multi-organ segmentation is very common while challenging in the field of medical image analysis, where organs largely vary with scales. Different organs are often treated equally in most segmentation networks, which is not quite optimal. In this work, we propose to divide a multi-organ segmentation task into multiple binary segmentation tasks by constructing a multi-to-binary network (MTBNet). The proposed MTBNet is based on the FCN for pixel-wise prediction. Moreover, we build a plug-and-play multi-to-binary block (MTB block) to adjust the influence of the feature maps from the backbone. This is achieved by parallelizing multiple branches with different convolutional layers and a probability gate (ProbGate). The ProbGate is set up for predicting whether the class exists, which is supervised clearly via an auxiliary loss without using any other inputs. More reasonable features are achieved by summing branches' features multiplied by the probability from the accompanying ProbGate and fed into a decoder module for prediction. The proposed method is validated on a challenging task dataset of multi-organ segmentation in abdominal MRI. Compared to classic medical and other multi-scale segmentation methods, the proposed MTBNet improves the segmentation accuracy of small organs by adjusting features from different organs and reducing the chance of missing or misidentifying these organs.

Dosimetric predictors of patient-reported toxicity after prostate stereotactic body radiotherapy: Analysis of full range of the dose-volume histogram using ensemble machine learning.

BACKGROUND AND PURPOSE: This study aims to evaluate the associations between dosimetric parameters and patient-reported outcomes, and to identify latent dosimetric parameters that most correlate with acute and subacute patient-reported urinary and rectal toxicity after prostate stereotactic body radiotherapy (SBRT) using machine learning methods. MATERIALS AND METHODS: Eighty-six patients who underwent prostate SBRT (40 Gy in 5 fractions) were included. Patient-reported health-related quality of life (HRQOL) outcomes were derived from bowel and bladder symptom scores on the Expanded Prostate Cancer Index Composite (EPIC-26) at 3 and 12 months post-SBRT. We utilized ensemble machine learning (ML) to interrogate the entire dose-volume histogram (DVH) to evaluate relationships between dose-volume parameters and HRQOL changes. The latent predictive dosimetric parameters that were most associated with HRQOL changes in urinary and rectal function were thus identified. An external cohort of 26 prostate SBRT patients was acquired to further test the predictive models. RESULTS: Bladder dose-volume metrics strongly predicted patient-reported urinary irritative and incontinence symptoms (area under the curves [AUCs] of 0.79 and 0.87, respectively) at 12 months. Maximum bladder dose, bladder V102.5%, bladder volume, and conformity indices (V50/VPTV and V100/VPTV) were most predictive of HRQOL changes in both urinary domains. No strong rectal toxicity dosimetric association was identified (AUC = 0.64). CONCLUSION: We demonstrated the application of advanced ML methods to identify a set of dosimetric variables that most highly correlated with patient-reported urinary HRQOL. DVH quantities identified with these methods may be used to achieve outcome-driven planning objectives to further reduce patient-reported toxicity with prostate SBRT.

Deep learning-based radiomic features for improving neoadjuvant chemoradiation response prediction in locally advanced rectal cancer.

Radiomic features achieve promising results in cancer diagnosis, treatment response prediction, and survival prediction. Our goal is to compare the handcrafted (explicitly designed) and deep learning (DL)-based radiomic features extracted from pre-treatment diffusion-weighted magnetic resonance images (DWIs) for predicting neoadjuvant chemoradiation treatment (nCRT) response in patients with locally advanced rectal cancer (LARC). 43 Patients receiving nCRT were included. All patients underwent DWIs before nCRT and total mesorectal excision surgery 6-12 weeks after completion of nCRT. Gross tumor volume (GTV) contours were drawn by an experienced radiation oncologist on DWIs. The patient-cohort was split into the responder group (n = 22) and the non-responder group (n = 21) based on the post-nCRT response assessed by postoperative pathology, MRI or colonoscopy. Handcrafted and DL-based features were extracted from the apparent diffusion coefficient (ADC) map of the DWI using conventional computer-aided diagnosis methods and a pre-trained convolution neural network, respectively. Least absolute shrinkage and selection operator (LASSO)-logistic regression models were constructed using extracted features for predicting treatment response. The model performance was evaluated with repeated 20 times stratified 4-fold cross-validation using receiver operating characteristic (ROC) curves and compared using the corrected paired t-test. The model built with handcrafted features achieved the mean area under the ROC curve (AUC) of 0.64, while the one built with DL-based features yielded the mean AUC of 0.73. The corrected paired t-test on AUC showed P-value < 0.05. DL-based features extracted from pre-treatment DWIs achieved significantly better classification performance compared with handcrafted features for predicting nCRT response in patients with LARC.