Predicting radiotherapy outcomes using statistical learning techniques

Utilization of machine learning methods for prediction of acute and late rectal toxicity due to curative prostate radiotherapy

OBJECTIVE

Rectal toxicity is one of the primary dose-limiting side effects of prostate cancer radiotherapy, and consequential impairment on quality of life in these patients with long survival is an important problem. In this study, we aimed to evaluate the possibility of predicting rectal toxicity with artificial intelligence model which was including certain dosimetric parameters.

MATERIALS AND METHODS

One hundred and thirty-seven patients with a diagnosis of prostate cancer who received curative radiotherapy for prostate +/- pelvic lymphatics were included in the study. The association of the clinical data and dosimetric data between early and late rectal toxicity reported during follow-up was evaluated. The sample size was increased to 274 patients by synthetic data generation method. To determine suitable models, 15 models were studied with machine learning algorithms using Python 2.3, Pycaret library. Random forest classifier was used with to detect active variables.

RESULTS

The area under the curve and accuracy were found to be 0.89-0.97 and 95%-99%, respectively, with machine learning algorithms. The sensitivity values for acute and toxicity were found to be 0.95 and 0.99, respectively.

CONCLUSION

Early or late rectal toxicity can be predicted with a high probability via dosimetric and physical data and machine learning algorithms of patients who underwent prostate +/- pelvic radiotherapy. The fact that rectal toxicity can be predicted before treatment, which may result in limiting the dose and duration of treatment, makes us think that artificial intelligence can enter our daily practice in a short time in this sense.

Embedding machine learning based toxicity models within radiotherapy treatment plan optimization

Objective.This study addresses radiation-induced toxicity (RIT) challenges in radiotherapy (RT) by developing a personalized treatment planning framework. It leverages patient-specific data and dosimetric information to create an optimization model that limits adverse side effects using constraints learned from historical data.Approach.The study uses the optimization with constraint learning (OCL) framework, incorporating patient-specific factors into the optimization process. It consists of three steps: optimizing the baseline treatment plan using population-wide dosimetric constraints; training a machine learning (ML) model to estimate the patient's RIT for the baseline plan; and adapting the treatment plan to minimize RIT using ML-learned patient-specific constraints. Various predictive models, including classification trees, ensembles of trees, and neural networks, are applied to predict the probability of grade 2+ radiation pneumonitis (RP2+) for non-small cell lung (NSCLC) cancer patients three months post-RT. The methodology is assessed with four high RP2+ risk NSCLC patients, with the goal of optimizing the dose distribution to constrain the RP2+ outcome below a pre-specified threshold. Conventional and OCL-enhanced plans are compared based on dosimetric parameters and predicted RP2+ risk. Sensitivity analysis on risk thresholds and data uncertainty is performed using a toy NSCLC case.Main results.Experiments show the methodology's capacity to directly incorporate all predictive models into RT treatment planning. In the four patients studied, mean lung dose and V20 were reduced by an average of 1.78 Gy and 3.66%, resulting in an average RP2+ risk reduction from 95% to 42%. Notably, this reduction maintains tumor coverage, although in two cases, sparing the lung slightly increased spinal cord max-dose (0.23 and 0.79 Gy).Significance.By integrating patient-specific information into learned constraints, the study significantly reduces adverse side effects like RP2+ without compromising target coverage. This unified framework bridges the gap between predicting toxicities and optimizing treatment plans in personalized RT decision-making.

Artificial intelligence to predict outcomes of head and neck radiotherapy

Head and neck radiotherapy induces important toxicity, and its efficacy and tolerance vary widely across patients. Advancements in radiotherapy delivery techniques, along with the increased quality and frequency of image guidance, offer a unique opportunity to individualize radiotherapy based on imaging biomarkers, with the aim of improving radiation efficacy while reducing its toxicity. Various artificial intelligence models integrating clinical data and radiomics have shown encouraging results for toxicity and cancer control outcomes prediction in head and neck cancer radiotherapy. Clinical implementation of these models could lead to individualized risk-based therapeutic decision making, but the reliability of the current studies is limited. Understanding, validating and expanding these models to larger multi-institutional data sets and testing them in the context of clinical trials is needed to ensure safe clinical implementation. This review summarizes the current state of the art of machine learning models for prediction of head and neck cancer radiotherapy outcomes.

Machine learning-based predictions of gamma passing rates for virtual specific-plan verification based on modulation maps, monitor unit profiles, and composite dose images

Machine learning (ML) methods have been implemented in radiotherapy to aid virtual specific-plan verification protocols, predicting gamma passing rates (GPR) based on calculated modulation complexity metrics because of their direct relation to dose deliverability. Nevertheless, these metrics might not comprehensively represent the modulation complexity, and automatically extracted features from alternative predictors associated with modulation complexity are needed. For this reason, three convolutional neural networks (CNN) based models were trained to predict GPR values (regression and classification), using respectively three predictors: (1) the modulation maps (MM) from the multi-leaf collimator, (2) the relative monitor units per control point profile (MUcp), and (3) the composite dose image (CDI) used for portal dosimetry, from 1024 anonymized prostate plans. The models' performance was assessed for classification and regression by the area under the receiver operator characteristic curve (AUC_ROC) and Spearman's correlation coefficient (r). Finally, four hybrid models were designed using all possible combinations of the three predictors. The prediction performance for the CNN-models using single predictors (MM, MUcp, and CDI) were AUC_ROC = 0.84 ± 0.03, 0.77 ± 0.07, 0.75 ± 0.04, andr= 0.6, 0.5, 0.7. Contrastingly, the hybrid models (MM + MUcp, MM + CDI, MUcp+CDI, MM + MUcp+CDI) performance were AUC_ROC = 0.94 ± 0.03, 0.85 ± 0.06, 0.89 ± 0.06, 0.91 ± 0.03, andr= 0.7, 0.5, 0.6, 0.7. The MP, MUcp, and CDI are suitable predictors for dose deliverability models implementing ML methods. Additionally, hybrid models are susceptible to improving their prediction performance, including two or more input predictors.

Predictive performance of different NTCP techniques for radiation-induced esophagitis in NSCLC patients receiving proton radiotherapy

This study aimed to compare the predictive performance of different modeling methods in developing normal tissue complication probability (NTCP) models for predicting radiation-induced esophagitis (RE) in non–small cell lung cancer (NSCLC) patients receiving proton radiotherapy. The dataset was composed of 328 NSCLC patients receiving passive-scattering proton therapy and 41.6% of the patients experienced ≥ grade 2 RE. Five modeling methods were used to build NTCP models: standard Lyman–Kutcher–Burman (sLKB), generalized LKB (gLKB), multivariable logistic regression using two variable selection procedures-stepwise forward selection (Stepwise-MLR), and least absolute shrinkage and selection operator (LASSO-MLR), and support vector machines (SVM). Predictive performance was internally validated by a bootstrap approach for each modeling method. The overall performance, discriminative ability, and calibration were assessed using the Negelkerke R², area under the receiver operator curve (AUC), and Hosmer–Lemeshow test, respectively. The LASSO-MLR model showed the best discriminative ability with an AUC value of 0.799 (95% confidence interval (CI): 0.763–0.854), and the best overall performance with a Negelkerke R² value of 0.332 (95% CI: 0.266–0.486). Both of the optimism-corrected Negelkerke R² values of the SVM and sLKB models were 0.301. The optimism-corrected AUC of the gLKB model (0.796) was higher than that of the SVM model (0.784). The sLKB model had the smallest optimism in the model variation and discriminative ability. In the context of classification and probability estimation for predicting the NTCP for radiation-induced esophagitis, the MLR model developed with LASSO provided the best predictive results. The simplest LKB modeling had similar or even better predictive performance than the most complex SVM modeling, and it was least likely to overfit the training data. The advanced machine learning approach might have limited applicability in clinical settings with a relatively small amount of data.

Integrating Multiomics Information in Deep Learning Architectures for Joint Actuarial Outcome Prediction in Non-Small Cell Lung Cancer Patients After Radiation Therapy

PURPOSE

Novel actuarial deep learning neural network (ADNN) architectures are proposed for joint prediction of radiation therapy outcomes-radiation pneumonitis (RP) and local control (LC)-in stage III non-small cell lung cancer (NSCLC) patients. Unlike normal tissue complication probability/tumor control probability models that use dosimetric information solely, our proposed models consider complex interactions among multiomics information including positron emission tomography (PET) radiomics, cytokines, and miRNAs. Additional time-to-event information is also used in the actuarial prediction.

METHODS AND MATERIALS

Three architectures were investigated: ADNN-DVH considered dosimetric information only; ADNN-com integrated multiomics information; and ADNN-com-joint combined RP2 (RP grade ≥2) and LC prediction. In these architectures, differential dose-volume histograms (DVHs) were fed into 1D convolutional neural networks (CNN) for extracting reduced representations. Variational encoders were used to learn representations of imaging and biological data. Reduced representations were fed into Surv-Nets to predict time-to-event probabilities for RP2 and LC independently and jointly by incorporating time information into designated loss functions.

RESULTS

Models were evaluated on 117 retrospective patients and were independently tested on 25 newly accrued patients prospectively. A multi-institutional RTOG0617 data set of 327 patients was used for external validation. ADNN-DVH yielded cross-validated c-indexes (95% confidence intervals) of 0.660 (0.630-0.690) for RP2 prediction and 0.727 (0.700-0.753) for LC prediction, outperforming a generalized Lyman model for RP2 (0.613 [0.583-0.643]) and a generalized log-logistic model for LC (0.569 [0.545-0.594]). The independent internal test and external validation yielded similar results. ADNN-com achieved an even better performance than ADNN-DVH on both cross-validation and independent internal test. Furthermore, ADNN-com-joint, which yielded performance similar to ADNN-com, realized joint prediction with c-indexes of 0.705 (0.676-0.734) for RP2 and 0.740 (0.714-0.765) for LC and achieved an area under a free-response receiving operator characteristic curve (AU-FROC) of 0.729 (0.697-0.773) for the joint prediction of RP2 and LC.

CONCLUSION

Novel deep learning architectures that integrate multiomics information outperformed traditional normal tissue complication probability/tumor control probability models in actuarial prediction of RP2 and LC.

Machine learning applications in radiation oncology

Machine learning technology has a growing impact on radiation oncology with an increasing presence in research and industry. The prevalence of diverse data including 3D imaging and the 3D radiation dose delivery presents potential for future automation and scope for treatment improvements for cancer patients. Harnessing this potential requires standardization of tools and data, and focused collaboration between fields of expertise. The rapid advancement of radiation oncology treatment technologies presents opportunities for machine learning integration with investments targeted towards data quality, data extraction, software, and engagement with clinical expertise. In this review, we provide an overview of machine learning concepts before reviewing advances in applying machine learning to radiation oncology and integrating these techniques into the radiation oncology workflows. Several key areas are outlined in the radiation oncology workflow where machine learning has been applied and where it can have a significant impact in terms of efficiency, consistency in treatment and overall treatment outcomes. This review highlights that machine learning has key early applications in radiation oncology due to the repetitive nature of many tasks that also currently have human review. Standardized data management of routinely collected imaging and radiation dose data are also highlighted as enabling engagement in research utilizing machine learning and the ability integrate these technologies into clinical workflow to benefit patients. Physicists need to be part of the conversation to facilitate this technical integration.

Spatial descriptions of radiotherapy dose: normal tissue complication models and statistical associations

For decades, dose-volume information for segmented anatomy has provided the essential data for correlating radiotherapy dosimetry with treatment-induced complications. Dose-volume information has formed the basis for modelling those associations via normal tissue complication probability (NTCP) models and for driving treatment planning. Limitations to this approach have been identified. Many studies have emerged demonstrating that the incorporation of information describing the spatial nature of the dose distribution, and potentially its correlation with anatomy, can provide more robust associations with toxicity and seed more general NTCP models. Such approaches are culminating in the application of computationally intensive processes such as machine learning and the application of neural networks. The opportunities these approaches have for individualising treatment, predicting toxicity and expanding the solution space for radiation therapy are substantial and have clearly widespread and disruptive potential. Impediments to reaching that potential include issues associated with data collection, model generalisation and validation. This review examines the role of spatial models of complication and summarises relevant published studies. Sources of data for these studies, appropriate statistical methodology frameworks for processing spatial dose information and extracting relevant features are described. Spatial complication modelling is consolidated as a pathway to guiding future developments towards effective, complication-free radiotherapy treatment.

Machine Learning to Predict Treatment in Oropharyngeal Squamous Cell Carcinoma

OBJECTIVES

This study seeks to (1) demonstrate how machine learning (ML) can be used for prediction modeling by predicting the treatment patients with T1-2, N0-N1 oropharyngeal squamous cell carcinoma (OPSCC) receive and (2) assess the impact patient, socioeconomic, regional, and institutional factors have in the treatment of this population.

METHODS

A retrospective cohort of adults diagnosed with T1-2, N0-N1 OPSCC from 2004 to 2013 was obtained using the National Cancer Database. The data was split into 80/20 distribution for training and testing, respectively. Various ML algorithms were explored for development. Area under the curve (AUC), accuracy, precision, and recall were calculated for the final model.

RESULTS

Among the 19,111 patients in the study, the mean (standard deviation) age was 61.3 (10.8) years, 14,034 (73%) were male, and 17,292 (91%) were white. Surgery was the primary treatment in 9,533 (50%) cases and radiation in 9,578 (50%) cases. The model heavily utilized T-stage, primary site, N-stage, grade, and type of treatment facility to predict the primary treatment modality. The final model yielded an AUC of 78% (95% CI, 77-79%), accuracy of 71%, precision of 72%, and recall of 71%.

CONCLUSION

This study created a ML model utilizing clinical variables to predict primary treatment modality for T1-2, N0-N1 OPSCC. This study demonstrates how ML can be used for prediction modeling while also highlighting that tumor and facility realted variables impact the decision making process on a national level.

Targeted transfer learning to improve performance in small medical physics datasets

PURPOSE

To perform an in-depth evaluation of current state of the art techniques in training neural networks to identify appropriate approaches in small datasets.

METHOD

In total, 112,120 frontal-view X-ray images from the NIH ChestXray14 dataset were used in our analysis. Two tasks were studied: unbalanced multi-label classification of 14 diseases, and binary classification of pneumonia vs non-pneumonia. All datasets were randomly split into training, validation, and testing (70%, 10%, and 20%). Two popular convolution neural networks (CNNs), DensNet121 and ResNet50, were trained using PyTorch. We performed several experiments to test: (a) whether transfer learning using pretrained networks on ImageNet are of value to medical imaging/physics tasks (e.g., predicting toxicity from radiographic images after training on images from the internet), (b) whether using pretrained networks trained on problems that are similar to the target task helps transfer learning (e.g., using X-ray pretrained networks for X-ray target tasks), (c) whether freeze deep layers or change all weights provides an optimal transfer learning strategy, (d) the best strategy for the learning rate policy, and (e) what quantity of data is needed in order to appropriately deploy these various strategies (N = 50 to N = 77 880).

RESULTS

In the multi-label problem, DensNet121 needed at least 1600 patients to be comparable to, and 10 000 to outperform, radiomics-based logistic regression. In classifying pneumonia vs non-pneumonia, both CNN and radiomics-based methods performed poorly when N < 2000. For small datasets ( < 2000), however, a significant boost in performance (>15% increase on AUC) comes from a good selection of the transfer learning dataset, dropout, cycling learning rate, and freezing and unfreezing of deep layers as training progresses. In contrast, if sufficient data are available (>35 000), little or no tweaking is needed to obtain impressive performance. While transfer learning using X-ray images from other anatomical sites improves performance, we also observed a similar boost by using pretrained networks from ImageNet. Having source images from the same anatomical site, however, outperforms every other methodology, by up to 15%. In this case, DL models can be trained with as little as N = 50.

CONCLUSIONS

While training DL models in small datasets (N < 2000) is challenging, no tweaking is necessary for bigger datasets (N > 35 000). Using transfer learning with images from the same anatomical site can yield remarkable performance in new tasks with as few as N = 50. Surprisingly, we did not find any advantage for using images from other anatomical sites over networks that have been trained using ImageNet. This indicates that features learned may not be as general as currently believed, and performance decays rapidly even by just changing the anatomical site of the images.

Machine Learning-Based Models for Prediction of Toxicity Outcomes in Radiotherapy

In order to limit radiotherapy (RT)-related side effects, effective toxicity prediction and assessment schemes are essential. In recent years, the growing interest toward artificial intelligence and machine learning (ML) within the science community has led to the implementation of innovative tools in RT. Several researchers have demonstrated the high performance of ML-based models in predicting toxicity, but the application of these approaches in clinics is still lagging, partly due to their low interpretability. Therefore, an overview of contemporary research is needed in order to familiarize practitioners with common methods and strategies. Here, we present a review of ML-based models for predicting and classifying RT-induced complications from both a methodological and a clinical standpoint, focusing on the type of features considered, the ML methods used, and the main results achieved. Our work overviews published research in multiple cancer sites, including brain, breast, esophagus, gynecological, head and neck, liver, lung, and prostate cancers. The aim is to define the current state of the art and main achievements within the field for both researchers and clinicians.

Evaluation of a neural network‐based photon beam profile deconvolution method

Purpose

The authors have previously shown the feasibility of using an artificial neural network (ANN) to eliminate the volume average effect (VAE) of scanning ionization chambers (ICs). The purpose of this work was to evaluate the method when applied to beams of different energies (6 and 10 MV) and modalities [flattened (FF) vs unflattened (FFF)], measured with ICs of various sizes.

Methods

The three‐layer ANN extracted data from transverse photon beam profiles using a sliding window, and output deconvolved value corresponding to the location at the center of the window. Beam profiles of seven fields ranging from 2 × 2 to 10 × 10 cm² at four depths (1.5, 5, 10 and 20 cm) were measured with three ICs (CC04, CC13, and FC65‐P) and an EDGE diode detector for 6 MV FF and FFF. Similar data for the 10 MV FF beam was also collected with CC13 and EDGE. The EDGE‐measured profiles were used as reference data to train and test the ANNs. Separate ANNs were trained by using the data of each beam energy and modality. Combined ANNs were also trained by combining data of different beam energies and/or modalities. The ANN's performance was quantified and compared by evaluating the penumbra width difference (PWD) between the deconvolved and reference profiles.

Results

Excellent agreement between the deconvolved and reference profiles was achieved with both separate and combined ANNs for all studied ICs, beam energies, beam modalities, and geometries. After deconvolution, the average PWD decreased from 1–3 mm to under 0.15 mm with separate ANNs and to under 0.20 mm with combined ANN.

Conclusions

The ANN‐based deconvolution method can be effectively applied to beams of different energies and modalities measured with ICs of various sizes. Separate ANNs yielded marginally better results than combined ANNs. An IC‐specific, combined ANN can provide clinically acceptable results as long as the training data includes data of each beam energy and modality.

Artificial Intelligence: reshaping the practice of radiological sciences in the 21st century

Advances in computing hardware and software platforms have led to the recent resurgence in artificial intelligence (AI) touching almost every aspect of our daily lives by its capability for automating complex tasks or providing superior predictive analytics. AI applications are currently spanning many diverse fields from economics to entertainment, to manufacturing, as well as medicine. Since modern AI's inception decades ago, practitioners in radiological sciences have been pioneering its development and implementation in medicine, particularly in areas related to diagnostic imaging and therapy. In this anniversary article, we embark on a journey to reflect on the learned lessons from past AI's chequered history. We further summarize the current status of AI in radiological sciences, highlighting, with examples, its impressive achievements and effect on re-shaping the practice of medical imaging and radiotherapy in the areas of computer-aided detection, diagnosis, prognosis, and decision support. Moving beyond the commercial hype of AI into reality, we discuss the current challenges to overcome, for AI to achieve its promised hope of providing better precision healthcare for each patient while reducing cost burden on their families and the society at large.

Optimization of treatment strategy by using a machine learning model to predict survival time of patients with malignant glioma after radiotherapy

The purpose of this study was to predict the survival time of patients with malignant glioma after radiotherapy with high accuracy by considering additional clinical factors and optimize the prescription dose and treatment duration for individual patient by using a machine learning model. A total of 35 patients with malignant glioma were included in this study. The candidate features included 12 clinical features and 192 dose-volume histogram (DVH) features. The appropriate input features and parameters of the support vector machine (SVM) were selected using the genetic algorithm based on Akaike's information criterion, i.e. clinical, DVH, and both clinical and DVH features. The prediction accuracy of the SVM models was evaluated through a leave-one-out cross-validation test with residual error, which was defined as the absolute difference between the actual and predicted survival times after radiotherapy. Moreover, the influences of various values of prescription dose and treatment duration on the predicted survival time were evaluated. The prediction accuracy was significantly improved with the combined use of clinical and DVH features compared with the separate use of both features (P < 0.01, Wilcoxon signed rank test). Mean ± standard deviation of the leave-one-out cross-validation using the combined clinical and DVH features, only clinical features and only DVH features were 104.7 ± 96.5, 144.2 ± 126.1 and 204.5 ± 186.0 days, respectively. The prediction accuracy could be improved with the combination of clinical and DVH features, and our results show the potential to optimize the treatment strategy for individual patients based on a machine learning model.

Combining handcrafted features with latent variables in machine learning for prediction of radiation-induced lung damage

PURPOSE

There has been burgeoning interest in applying machine learning methods for predicting radiotherapy outcomes. However, the imbalanced ratio of a large number of variables to a limited sample size in radiation oncology constitutes a major challenge. Therefore, dimensionality reduction methods can be a key to success. The study investigates and contrasts the application of traditional machine learning methods and deep learning approaches for outcome modeling in radiotherapy. In particular, new joint architectures based on variational autoencoder (VAE) for dimensionality reduction are presented and their application is demonstrated for the prediction of lung radiation pneumonitis (RP) from a large-scale heterogeneous dataset.

METHODS

A large-scale heterogeneous dataset containing a pool of 230 variables including clinical factors (e.g., dose, KPS, stage) and biomarkers (e.g., single nucleotide polymorphisms (SNPs), cytokines, and micro-RNAs) in a population of 106 nonsmall cell lung cancer (NSCLC) patients who received radiotherapy was used for modeling RP. Twenty-two patients had grade 2 or higher RP. Four methods were investigated, including feature selection (case A) and feature extraction (case B) with traditional machine learning methods, a VAE-MLP joint architecture (case C) with deep learning and lastly, the combination of feature selection and joint architecture (case D). For feature selection, Random forest (RF), Support Vector Machine (SVM), and multilayer perceptron (MLP) were implemented to select relevant features. Specifically, each method was run for multiple times to rank features within several cross-validated (CV) resampled sets. A collection of ranking lists were then aggregated by top 5% and Kemeny graph methods to identify the final ranking for prediction. A synthetic minority oversampling technique was applied to correct for class imbalance during this process. For deep learning, a VAE-MLP joint architecture where a VAE aimed for dimensionality reduction and an MLP aimed for classification was developed. In this architecture, reconstruction loss and prediction loss were combined into a single loss function to realize simultaneous training and weights were assigned to different classes to mitigate class imbalance. To evaluate the prediction performance and conduct comparisons, the area under receiver operating characteristic curves (AUCs) were performed for nested CVs for both handcrafted feature selections and the deep learning approach. The significance of differences in AUCs was assessed using the DeLong test of U-statistics.

RESULTS

An MLP-based method using weight pruning (WP) feature selection yielded the best performance among the different hand-crafted feature selection methods (case A), reaching an AUC of 0.804 (95% CI: 0.761-0.823) with 29 top features. A VAE-MLP joint architecture (case C) achieved a comparable but slightly lower AUC of 0.781 (95% CI: 0.737-0.808) with the size of latent dimension being 2. The combination of handcrafted features (case A) and latent representation (case D) achieved a significant AUC improvement of 0.831 (95% CI: 0.805-0.863) with 22 features (P-value = 0.000642 compared with handcrafted features only (Case A) and P-value = 0.000453 compared to VAE alone (Case C)) with an MLP classifier.

CONCLUSION

The potential for combination of traditional machine learning methods and deep learning VAE techniques has been demonstrated for dealing with limited datasets in modeling radiotherapy toxicities. Specifically, latent variables from a VAE-MLP joint architecture are able to complement handcrafted features for the prediction of RP and improve prediction over either method alone.

Predicting radiation pneumonitis in locally advanced stage II-III non-small cell lung cancer using machine learning

BACKGROUND AND PURPOSE

Radiation pneumonitis (RP) is a radiotherapy dose-limiting toxicity for locally advanced non-small cell lung cancer (LA-NSCLC). Prior studies have proposed relevant dosimetric constraints to limit this toxicity. Using machine learning algorithms, we performed analyses of contributing factors in the development of RP to uncover previously unidentified criteria and elucidate the relative importance of individual factors.

MATERIALS AND METHODS

We evaluated 32 clinical features per patient in a cohort of 203 stage II-III LA-NSCLC patients treated with definitive chemoradiation to a median dose of 66.6 Gy in 1.8 Gy daily fractions at our institution from 2008 to 2016. Of this cohort, 17.7% of patients developed grade ≥2 RP. Univariate analysis was performed using trained decision stumps to individually analyze statistically significant predictors of RP and perform feature selection. Applying Random Forest, we performed multivariate analysis to assess the combined performance of important predictors of RP.

RESULTS

On univariate analysis, lung V20, lung mean, lung V10 and lung V5 were found to be significant RP predictors with the greatest balance of specificity and sensitivity. On multivariate analysis, Random Forest (AUC = 0.66, p = 0.0005) identified esophagus max (20.5%), lung V20 (16.4%), lung mean (15.7%) and pack-year (14.9%) as the most common primary differentiators of RP.

CONCLUSIONS

We highlight Random Forest as an accurate machine learning method to identify known and new predictors of symptomatic RP. Furthermore, this analysis confirms the importance of lung V20, lung mean and pack-year as predictors of RP while also introducing esophagus max as an important RP predictor.

Prospects and challenges for clinical decision support in the era of big data

Recently, there has been burgeoning interest in developing more effective and robust clinical decision support systems (CDSSs) for oncology. This has been primarily driven by the demands for more personalized and precise medical practice in oncology in the era of so-called Big Data (BD); an era that promises to harness the power of large-scale data flow to revolutionize cancer treatment. This interest in BD analytics has created new opportunities as well as new unmet challenges. These include: routine aggregation and standardization of clinical data; patient privacy; transformation of current analytical approaches to handle such noisy and heterogeneous data; and expanded use of advanced statistical learning methods based on confluence of modern statistical methods and machine learning algorithms. In this review, we present the current status of CDSSs in oncology, the prospects and current challenges of BD analytics, and the promising role of integrated modern statistics and machine learning algorithms in predicting complex clinical endpoints, individualizing treatment rules, and optimizing dynamic personalized treatment regimens. We discuss issues pertaining to these topics and present application examples from an aggregate of experiences. We also discuss the role of human factors in improving the utilization and acceptance of such enhanced CDSSs and how to mitigate possible sources of human error to achieve optimal performance and wider acceptance.

DW-MRI and DCE-MRI are of complementary value in predicting pathologic response to neoadjuvant chemoradiotherapy for esophageal cancer

PURPOSE

To explore the potential benefit and complementary value of a multiparametric approach using diffusion-weighted (DW-) and dynamic contrast-enhanced (DCE-) magnetic resonance imaging (MRI) for prediction of response to neoadjuvant chemoradiotherapy (nCRT) in esophageal cancer.

MATERIAL AND METHODS

Forty-five patients underwent both DW-MRI and DCE-MRI prior to nCRT (pre), during nCRT (week 2-3) (per) and after completion of nCRT, but prior to esophagectomy (post). Subsequently, histopathologic tumor regression grade (TRG) was assessed. Tumor apparent diffusion coefficient (ADC) and area-under-the-concentration time curve (AUC) were calculated for DW-MRI and DCE-MRI, respectively. The ability of these parameters to predict pathologic complete response (pCR, TRG1) or good response (GR, TRG ≤ 2) to nCRT was assessed. Furthermore the complementary value of DW-MRI and DCE-MRI was investigated.

RESULTS

GR was found in 22 (49%) patients, of which 10 (22%) patients showed pCR. For DW-MRI, the 75th percentile (P75) ΔADC_post-pre was most predictive for GR (c-index = 0.75). For DCE-MRI, P90 ΔAUC_per-pre was most predictive for pCR (c-index = 0.79). Multivariable logistic regression analyses showed complementary value when combining DW-MRI and DCE-MRI for pCR prediction (c-index = 0.89).

CONCLUSIONS

Both DW-MRI and DCE-MRI are promising in predicting response to nCRT in esophageal cancer. Combining both modalities provides complementary information, resulting in a higher predictive value.

Machine learning and modeling: Data, validation, communication challenges

With the era of big data, the utilization of machine learning algorithms in radiation oncology is rapidly growing with applications including: treatment response modeling, treatment planning, contouring, organ segmentation, image-guidance, motion tracking, quality assurance, and more. Despite this interest, practical clinical implementation of machine learning as part of the day-to-day clinical operations is still lagging. The aim of this white paper is to further promote progress in this new field of machine learning in radiation oncology by highlighting its untapped advantages and potentials for clinical advancement, while also presenting current challenges and open questions for future research. The targeted audience of this paper includes newcomers as well as practitioners in the field of medical physics/radiation oncology. The paper also provides general recommendations to avoid common pitfalls when applying these powerful data analytic tools to medical physics and radiation oncology problems and suggests some guidelines for transparent and informative reporting of machine learning results.

Development of a Ready-to-Use Graphical Tool Based on Artificial Neural Network Classification: Application for the Prediction of Late Fecal Incontinence After Prostate Cancer Radiation Therapy

PURPOSE

This study was designed to apply artificial neural network (ANN) classification methods for the prediction of late fecal incontinence (LFI) after high-dose prostate cancer radiation therapy and to develop a ready-to-use graphical tool.

MATERIALS AND METHODS

In this study, 598 men recruited in 2 national multicenter trials were analyzed. Information was recorded on comorbidity, previous abdominal surgery, use of drugs, and dose distribution. Fecal incontinence was prospectively evaluated through self-reported questionnaires. To develop the ANN, the study population was randomly split into training (n = 300), validation (n = 149), and test (n = 149) sets. Mean grade of longitudinal LFI (ie, expressed as the average incontinence grade over the first 3 years after radiation therapy) ≥1 was considered the endpoint. A suitable subset of variables able to better predict LFI was selected by simulating 100,000 ANN configurations. The search for the definitive ANN was then performed by varying the number of inputs and hidden neurons from 4 to 5 and from 1 to 9, respectively. A final classification model was established as the average of the best 5 among 500 ANNs with the same architecture. An ANN-based graphical method to compute LFI prediction was developed to include one continuous and n dichotomous variables.

RESULTS

An ANN architecture was selected, with 5 input variables (mean dose, previous abdominal surgery, use of anticoagulants, use of antihypertensive drugs, and use of neoadjuvant and adjuvant hormone therapy) and 4 hidden neurons. The developed classification model correctly identified patients with LFI with 80.8% sensitivity and 63.7% ± 1.0% specificity and an area under the curve of 0.78. The developed graphical tool may efficiently classify patients in low, intermediate, and high LFI risk classes.

CONCLUSIONS

An ANN-based model was developed to predict LFI. The model was translated in a ready-to-use graphical tool for LFI risk classification, with direct interpretation of the role of the predictors.

Data Analysis in Radiotherapy Treatments

Radiotherapy is one of the main cancer treatments available today, together with chemotherapy and surgery. Radiotherapy treatments have to be planned for each patient in an individualized manner. The knowledge acquired from one single treatment can be used to improve the treatment planning and outcome of several other patients. In the last years, attention has been drawn to the added value of using data analysis for radiotherapy treatment planning, prediction of treatment outcomes, survival analysis and quality assurance. In this article, existing literature is reviewed.

Machine Learning and Radiogenomics: Lessons Learned and Future Directions

Due to the rapid increase in the availability of patient data, there is significant interest in precision medicine that could facilitate the development of a personalized treatment plan for each patient on an individual basis. Radiation oncology is particularly suited for predictive machine learning (ML) models due to the enormous amount of diagnostic data used as input and therapeutic data generated as output. An emerging field in precision radiation oncology that can take advantage of ML approaches is radiogenomics, which is the study of the impact of genomic variations on the sensitivity of normal and tumor tissue to radiation. Currently, patients undergoing radiotherapy are treated using uniform dose constraints specific to the tumor and surrounding normal tissues. This is suboptimal in many ways. First, the dose that can be delivered to the target volume may be insufficient for control but is constrained by the surrounding normal tissue, as dose escalation can lead to significant morbidity and rare. Second, two patients with nearly identical dose distributions can have substantially different acute and late toxicities, resulting in lengthy treatment breaks and suboptimal control, or chronic morbidities leading to poor quality of life. Despite significant advances in radiogenomics, the magnitude of the genetic contribution to radiation response far exceeds our current understanding of individual risk variants. In the field of genomics, ML methods are being used to extract harder-to-detect knowledge, but these methods have yet to fully penetrate radiogenomics. Hence, the goal of this publication is to provide an overview of ML as it applies to radiogenomics. We begin with a brief history of radiogenomics and its relationship to precision medicine. We then introduce ML and compare it to statistical hypothesis testing to reflect on shared lessons and to avoid common pitfalls. Current ML approaches to genome-wide association studies are examined. The application of ML specifically to radiogenomics is next presented. We end with important lessons for the proper integration of ML into radiogenomics.

Design and Selection of Machine Learning Methods Using Radiomics and Dosiomics for Normal Tissue Complication Probability Modeling of Xerostomia

Purpose

The purpose of this study is to investigate whether machine learning with dosiomic, radiomic, and demographic features allows for xerostomia risk assessment more precise than normal tissue complication probability (NTCP) models based on the mean radiation dose to parotid glands.

Material and methods

A cohort of 153 head-and-neck cancer patients was used to model xerostomia at 0–6 months (early), 6–15 months (late), 15–24 months (long-term), and at any time (a longitudinal model) after radiotherapy. Predictive power of the features was evaluated by the area under the receiver operating characteristic curve (AUC) of univariate logistic regression models. The multivariate NTCP models were tuned and tested with single and nested cross-validation, respectively. We compared predictive performance of seven classification algorithms, six feature selection methods, and ten data cleaning/class balancing techniques using the Friedman test and the Nemenyi post hoc analysis.

Results

NTCP models based on the parotid mean dose failed to predict xerostomia (AUCs < 0.60). The most informative predictors were found for late and long-term xerostomia. Late xerostomia correlated with the contralateral dose gradient in the anterior–posterior (AUC = 0.72) and the right–left (AUC = 0.68) direction, whereas long-term xerostomia was associated with parotid volumes (AUCs > 0.85), dose gradients in the right–left (AUCs > 0.78), and the anterior–posterior (AUCs > 0.72) direction. Multivariate models of long-term xerostomia were typically based on the parotid volume, the parotid eccentricity, and the dose–volume histogram (DVH) spread with the generalization AUCs ranging from 0.74 to 0.88. On average, support vector machines and extra-trees were the top performing classifiers, whereas the algorithms based on logistic regression were the best choice for feature selection. We found no advantage in using data cleaning or class balancing methods.

Conclusion

We demonstrated that incorporation of organ- and dose-shape descriptors is beneficial for xerostomia prediction in highly conformal radiotherapy treatments. Due to strong reliance on patient-specific, dose-independent factors, our results underscore the need for development of personalized data-driven risk profiles for NTCP models of xerostomia. The facilitated machine learning pipeline is described in detail and can serve as a valuable reference for future work in radiomic and dosiomic NTCP modeling.

Strategies for prediction and mitigation of radiation-induced liver toxicity

Although well described in the 1960s, liver toxicity secondary to radiation therapy, commonly known as radiation-induced liver disease (RILD), remains a major challenge. RILD encompasses two distinct clinical entities, a 'classic' form, composed of anicteric hepatomegaly, ascites and elevated alkaline phosphatase; and a 'non-classic' form, with liver transaminases elevated to more than five times the reference value, or worsening of liver metabolic function represented as an increase of 2 or more points in the Child-Pugh score classification. The risk of occurrence of RILD has historically limited the applicability of radiation for the treatment of liver malignancies. With the development of 3D conformal radiation therapy, which allowed for partial organ irradiation based on computed tomography treatment planning, there has been a resurgence of interest in the use of liver irradiation. Since then, a large body of evidence regarding the liver tolerance to conventionally fractionated radiation has been produced, but severe liver toxicities has continued to be reported. More recently, improvements in diagnostic imaging, radiation treatment planning technology and delivery systems have prompted the development of stereotactic body radiotherapy (SBRT), by which high doses of radiation can be delivered with high target accuracy and a steep dose gradient at the tumor - normal tissue interface, offering an opportunity of decreasing toxicity rates while improving tumor control. Here, we present an overview of the role SBRT has played in the management of liver tumors, addressing the challenges and opportunities to reduce the incidence of RILD, such as adaptive approaches and machine-learning-based predictive models.

Clinical decision support of radiotherapy treatment planning: A data-driven machine learning strategy for patient-specific dosimetric decision making

BACKGROUND AND PURPOSE

Clinical decision support systems are a growing class of tools with the potential to impact healthcare. This study investigates the construction of a decision support system through which clinicians can efficiently identify which previously approved historical treatment plans are achievable for a new patient to aid in selection of therapy.

MATERIAL AND METHODS

Treatment data were collected for early-stage lung and postoperative oropharyngeal cancers treated using photon (lung and head and neck) and proton (head and neck) radiotherapy. Machine-learning classifiers were constructed using patient-specific feature-sets and a library of historical plans. Model accuracy was analyzed using learning curves, and historical treatment plan matching was investigated.

RESULTS

Learning curves demonstrate that for these datasets, approximately 45, 60, and 30 patients are needed for a sufficiently accurate classification model for radiotherapy for early-stage lung, postoperative oropharyngeal photon, and postoperative oropharyngeal proton, respectively. The resulting classification model provides a database of previously approved treatment plans that are achievable for a new patient. An exemplary case, highlighting tradeoffs between the heart and chest wall dose while holding target dose constant in two historical plans is provided.

CONCLUSIONS

We report on the first artificial-intelligence based clinical decision support system that connects patients to past discrete treatment plans in radiation oncology and demonstrate for the first time how this tool can enable clinicians to use past decisions to help inform current assessments. Clinicians can be informed of dose tradeoffs between critical structures early in the treatment process, enabling more time spent on finding the optimal course of treatment for individual patients.

Radiogenomics and radiotherapy response modeling

Advances in patient-specific information and biotechnology have contributed to a new era of computational medicine. Radiogenomics has emerged as a new field that investigates the role of genetics in treatment response to radiation therapy. Radiation oncology is currently attempting to embrace these recent advances and add to its rich history by maintaining its prominent role as a quantitative leader in oncologic response modeling. Here, we provide an overview of radiogenomics starting with genotyping, data aggregation, and application of different modeling approaches based on modifying traditional radiobiological methods or application of advanced machine learning techniques. We highlight the current status and potential for this new field to reshape the landscape of outcome modeling in radiotherapy and drive future advances in computational oncology.

Statistical-learning strategies generate only modestly performing predictive models for urinary symptoms following external beam radiotherapy of the prostate: A comparison of conventional and machine-learning methods

PURPOSE

Given the paucity of available data concerning radiotherapy-induced urinary toxicity, it is important to ensure derivation of the most robust models with superior predictive performance. This work explores multiple statistical-learning strategies for prediction of urinary symptoms following external beam radiotherapy of the prostate.

METHODS

The performance of logistic regression, elastic-net, support-vector machine, random forest, neural network, and multivariate adaptive regression splines (MARS) to predict urinary symptoms was analyzed using data from 754 participants accrued by TROG03.04-RADAR. Predictive features included dose-surface data, comorbidities, and medication-intake. Four symptoms were analyzed: dysuria, haematuria, incontinence, and frequency, each with three definitions (grade ≥ 1, grade ≥ 2 and longitudinal) with event rate between 2.3% and 76.1%. Repeated cross-validations producing matched models were implemented. A synthetic minority oversampling technique was utilized in endpoints with rare events. Parameter optimization was performed on the training data. Area under the receiver operating characteristic curve (AUROC) was used to compare performance using sample size to detect differences of ≥0.05 at the 95% confidence level.

RESULTS

Logistic regression, elastic-net, random forest, MARS, and support-vector machine were the highest-performing statistical-learning strategies in 3, 3, 3, 2, and 1 endpoints, respectively. Logistic regression, MARS, elastic-net, random forest, neural network, and support-vector machine were the best, or were not significantly worse than the best, in 7, 7, 5, 5, 3, and 1 endpoints. The best-performing statistical model was for dysuria grade ≥ 1 with AUROC ± standard deviation of 0.649 ± 0.074 using MARS. For longitudinal frequency and dysuria grade ≥ 1, all strategies produced AUROC>0.6 while all haematuria endpoints and longitudinal incontinence models produced AUROC<0.6.

CONCLUSIONS

Logistic regression and MARS were most likely to be the best-performing strategy for the prediction of urinary symptoms with elastic-net and random forest producing competitive results. The predictive power of the models was modest and endpoint-dependent. New features, including spatial dose maps, may be necessary to achieve better models.

Predictive modelling analysis for development of a radiotherapy decision support system in prostate cancer: a preliminary study

Purpose

The aim of this study is to develop predictive models to predict organ at risk (OAR) complication level, classification of OAR dose-volume and combination of this function with our in-house developed treatment decision support system.

Materials and methods

We analysed the support vector machine and decision tree algorithm for predicting OAR complication level and toxicity in order to integrate this function into our in-house radiation treatment planning decision support system. A total of 12 TomoTherapyTM treatment plans for prostate cancer were established, and a hundred modelled plans were generated to analyse the toxicity prediction for bladder and rectum.

Results

The toxicity prediction algorithm analysis showed 91·0% accuracy in the training process. A scatter plot for bladder and rectum was obtained by 100 modelled plans and classification result derived. OAR complication level was analysed and risk factor for 25% bladder and 50% rectum was detected by decision tree. Therefore, it was shown that complication prediction of patients using big data-based clinical information is possible.

Conclusion

We verified the accuracy of the tested algorithm using prostate cancer cases. Side effects can be minimised by applying this predictive modelling algorithm with the planning decision support system for patient-specific radiotherapy planning.

Perspectives on making big data analytics work for oncology

Oncology, with its unique combination of clinical, physical, technological, and biological data provides an ideal case study for applying big data analytics to improve cancer treatment safety and outcomes. An oncology treatment course such as chemoradiotherapy can generate a large pool of information carrying the 5Vs hallmarks of big data. This data is comprised of a heterogeneous mixture of patient demographics, radiation/chemo dosimetry, multimodality imaging features, and biological markers generated over a treatment period that can span few days to several weeks. Efforts using commercial and in-house tools are underway to facilitate data aggregation, ontology creation, sharing, visualization and varying analytics in a secure environment. However, open questions related to proper data structure representation and effective analytics tools to support oncology decision-making need to be addressed. It is recognized that oncology data constitutes a mix of structured (tabulated) and unstructured (electronic documents) that need to be processed to facilitate searching and subsequent knowledge discovery from relational or NoSQL databases. In this context, methods based on advanced analytics and image feature extraction for oncology applications will be discussed. On the other hand, the classical p (variables)≫n (samples) inference problem of statistical learning is challenged in the Big data realm and this is particularly true for oncology applications where p-omics is witnessing exponential growth while the number of cancer incidences has generally plateaued over the past 5-years leading to a quasi-linear growth in samples per patient. Within the Big data paradigm, this kind of phenomenon may yield undesirable effects such as echo chamber anomalies, Yule-Simpson reversal paradox, or misleading ghost analytics. In this work, we will present these effects as they pertain to oncology and engage small thinking methodologies to counter these effects ranging from incorporating prior knowledge, using information-theoretic techniques to modern ensemble machine learning approaches or combination of these. We will particularly discuss the pros and cons of different approaches to improve mining of big data in oncology.

Using machine learning to predict radiation pneumonitis in patients with stage I non-small cell lung cancer treated with stereotactic body radiation therapy

To develop a patient-specific 'big data' clinical decision tool to predict pneumonitis in stage I non-small cell lung cancer (NSCLC) patients after stereotactic body radiation therapy (SBRT). 61 features were recorded for 201 consecutive patients with stage I NSCLC treated with SBRT, in whom 8 (4.0%) developed radiation pneumonitis. Pneumonitis thresholds were found for each feature individually using decision stumps. The performance of three different algorithms (Decision Trees, Random Forests, RUSBoost) was evaluated. Learning curves were developed and the training error analyzed and compared to the testing error in order to evaluate the factors needed to obtain a cross-validated error smaller than 0.1. These included the addition of new features, increasing the complexity of the algorithm and enlarging the sample size and number of events. In the univariate analysis, the most important feature selected was the diffusion capacity of the lung for carbon monoxide (DLCO adj%). On multivariate analysis, the three most important features selected were the dose to 15 cc of the heart, dose to 4 cc of the trachea or bronchus, and race. Higher accuracy could be achieved if the RUSBoost algorithm was used with regularization. To predict radiation pneumonitis within an error smaller than 10%, we estimate that a sample size of 800 patients is required. Clinically relevant thresholds that put patients at risk of developing radiation pneumonitis were determined in a cohort of 201 stage I NSCLC patients treated with SBRT. The consistency of these thresholds can provide radiation oncologists with an estimate of their reliability and may inform treatment planning and patient counseling. The accuracy of the classification is limited by the number of patients in the study and not by the features gathered or the complexity of the algorithm.

The value of nodal information in predicting lung cancer relapse using 4DPET/4DCT

PURPOSE

There is evidence that computed tomography (CT) and positron emission tomography (PET) imaging metrics are prognostic and predictive in nonsmall cell lung cancer (NSCLC) treatment outcomes. However, few studies have explored the use of standardized uptake value (SUV)-based image features of nodal regions as predictive features. The authors investigated and compared the use of tumor and node image features extracted from the radiotherapy target volumes to predict relapse in a cohort of NSCLC patients undergoing chemoradiation treatment.

METHODS

A prospective cohort of 25 patients with locally advanced NSCLC underwent 4DPET/4DCT imaging for radiation planning. Thirty-seven image features were derived from the CT-defined volumes and SUVs of the PET image from both the tumor and nodal target regions. The machine learning methods of logistic regression and repeated stratified five-fold cross-validation (CV) were used to predict local and overall relapses in 2 yr. The authors used well-known feature selection methods (Spearman's rank correlation, recursive feature elimination) within each fold of CV. Classifiers were ranked on their Matthew's correlation coefficient (MCC) after CV. Area under the curve, sensitivity, and specificity values are also presented.

RESULTS

For predicting local relapse, the best classifier found had a mean MCC of 0.07 and was composed of eight tumor features. For predicting overall relapse, the best classifier found had a mean MCC of 0.29 and was composed of a single feature: the volume greater than 0.5 times the maximum SUV (N).

CONCLUSIONS

The best classifier for predicting local relapse had only tumor features. In contrast, the best classifier for predicting overall relapse included a node feature. Overall, the methods showed that nodes add value in predicting overall relapse but not local relapse.

Automatic assessment of average diaphragm motion trajectory from 4DCT images through machine learning

To automatically estimate average diaphragm motion trajectory (ADMT) based on four-dimensional computed tomography (4DCT), facilitating clinical assessment of respiratory motion and motion variation and retrospective motion study. We have developed an effective motion extraction approach and a machine-learning-based algorithm to estimate the ADMT. Eleven patients with 22 sets of 4DCT images (4DCT1 at simulation and 4DCT2 at treatment) were studied. After automatically segmenting the lungs, the differential volume-per-slice (dVPS) curves of the left and right lungs were calculated as a function of slice number for each phase with respective to the full-exhalation. After 5-slice moving average was performed, the discrete cosine transform (DCT) was applied to analyze the dVPS curves in frequency domain. The dimensionality of the spectrum data was reduced by using several lowest frequency coefficients (f_v) to account for most of the spectrum energy (Σf_v²). Multiple linear regression (MLR) method was then applied to determine the weights of these frequencies by fitting the ground truth-the measured ADMT, which are represented by three pivot points of the diaphragm on each side. The 'leave-one-out' cross validation method was employed to analyze the statistical performance of the prediction results in three image sets: 4DCT1, 4DCT2, and 4DCT1 + 4DCT2. Seven lowest frequencies in DCT domain were found to be sufficient to approximate the patient dVPS curves (R = 91%-96% in MLR fitting). The mean error in the predicted ADMT using leave-one-out method was 0.3 ± 1.9 mm for the left-side diaphragm and 0.0 ± 1.4 mm for the right-side diaphragm. The prediction error is lower in 4DCT2 than 4DCT1, and is the lowest in 4DCT1 and 4DCT2 combined. This frequency-analysis-based machine learning technique was employed to predict the ADMT automatically with an acceptable error (0.2 ± 1.6 mm). This volumetric approach is not affected by the presence of the lung tumors, providing an automatic robust tool to evaluate diaphragm motion.

Machine Learning Approaches for Predicting Radiation Therapy Outcomes: A Clinician's Perspective

Radiation oncology has always been deeply rooted in modeling, from the early days of isoeffect curves to the contemporary Quantitative Analysis of Normal Tissue Effects in the Clinic (QUANTEC) initiative. In recent years, medical modeling for both prognostic and therapeutic purposes has exploded thanks to increasing availability of electronic data and genomics. One promising direction that medical modeling is moving toward is adopting the same machine learning methods used by companies such as Google and Facebook to combat disease. Broadly defined, machine learning is a branch of computer science that deals with making predictions from complex data through statistical models. These methods serve to uncover patterns in data and are actively used in areas such as speech recognition, handwriting recognition, face recognition, "spam" filtering (junk email), and targeted advertising. Although multiple radiation oncology research groups have shown the value of applied machine learning (ML), clinical adoption has been slow due to the high barrier to understanding these complex models by clinicians. Here, we present a review of the use of ML to predict radiation therapy outcomes from the clinician's point of view with the hope that it lowers the "barrier to entry" for those without formal training in ML. We begin by describing 7 principles that one should consider when evaluating (or creating) an ML model in radiation oncology. We next introduce 3 popular ML methods--logistic regression (LR), support vector machine (SVM), and artificial neural network (ANN)--and critique 3 seminal papers in the context of these principles. Although current studies are in exploratory stages, the overall methodology has progressively matured, and the field is ready for larger-scale further investigation.

Automated beam placement for breast radiotherapy using a support vector machine based algorithm

PURPOSE

To develop an automated beam placement technique for whole breast radiotherapy using tangential beams. We seek to find optimal parameters for tangential beams to cover the whole ipsilateral breast (WB) and minimize the dose to the organs at risk (OARs).

METHODS

A support vector machine (SVM) based method is proposed to determine the optimal posterior plane of the tangential beams. Relative significances of including/avoiding the volumes of interests are incorporated into the cost function of the SVM. After finding the optimal 3-D plane that separates the whole breast (WB) and the included clinical target volumes (CTVs) from the OARs, the gantry angle, collimator angle, and posterior jaw size of the tangential beams are derived from the separating plane equation. Dosimetric measures of the treatment plans determined by the automated method are compared with those obtained by applying manual beam placement by the physicians. The method can be further extended to use multileaf collimator (MLC) blocking by optimizing posterior MLC positions.

RESULTS

The plans for 36 patients (23 prone- and 13 supine-treated) with left breast cancer were analyzed. Our algorithm reduced the volume of the heart that receives >500 cGy dose (V5) from 2.7 to 1.7 cm(3) (p = 0.058) on average and the volume of the ipsilateral lung that receives >1000 cGy dose (V10) from 55.2 to 40.7 cm(3) (p = 0.0013). The dose coverage as measured by volume receiving >95% of the prescription dose (V95%) of the WB without a 5 mm superficial layer decreases by only 0.74% (p = 0.0002) and the V95% for the tumor bed with 1.5 cm margin remains unchanged.

CONCLUSIONS

This study has demonstrated the feasibility of using a SVM-based algorithm to determine optimal beam placement without a physician's intervention. The proposed method reduced the dose to OARs, especially for supine treated patients, without any relevant degradation of dose homogeneity and coverage in general.

Monte Carlo role in radiobiological modelling of radiotherapy outcomes

Radiobiological models are essential components of modern radiotherapy. They are increasingly applied to optimize and evaluate the quality of different treatment planning modalities. They are frequently used in designing new radiotherapy clinical trials by estimating the expected therapeutic ratio of new protocols. In radiobiology, the therapeutic ratio is estimated from the expected gain in tumour control probability (TCP) to the risk of normal tissue complication probability (NTCP). However, estimates of TCP/NTCP are currently based on the deterministic and simplistic linear-quadratic formalism with limited prediction power when applied prospectively. Given the complex and stochastic nature of the physical, chemical and biological interactions associated with spatial and temporal radiation induced effects in living tissues, it is conjectured that methods based on Monte Carlo (MC) analysis may provide better estimates of TCP/NTCP for radiotherapy treatment planning and trial design. Indeed, over the past few decades, methods based on MC have demonstrated superior performance for accurate simulation of radiation transport, tumour growth and particle track structures; however, successful application of modelling radiobiological response and outcomes in radiotherapy is still hampered with several challenges. In this review, we provide an overview of some of the main techniques used in radiobiological modelling for radiotherapy, with focus on the MC role as a promising computational vehicle. We highlight the current challenges, issues and future potentials of the MC approach towards a comprehensive systems-based framework in radiobiological modelling for radiotherapy.

Multivariate modeling of complications with data driven variable selection: guarding against overfitting and effects of data set size

PURPOSE

Multivariate modeling of complications after radiotherapy is frequently used in conjunction with data driven variable selection. This study quantifies the risk of overfitting in a data driven modeling method using bootstrapping for data with typical clinical characteristics, and estimates the minimum amount of data needed to obtain models with relatively high predictive power.

MATERIALS AND METHODS

To facilitate repeated modeling and cross-validation with independent datasets for the assessment of true predictive power, a method was developed to generate simulated data with statistical properties similar to real clinical data sets. Characteristics of three clinical data sets from radiotherapy treatment of head and neck cancer patients were used to simulate data with set sizes between 50 and 1000 patients. A logistic regression method using bootstrapping and forward variable selection was used for complication modeling, resulting for each simulated data set in a selected number of variables and an estimated predictive power. The true optimal number of variables and true predictive power were calculated using cross-validation with very large independent data sets.

RESULTS

For all simulated data set sizes the number of variables selected by the bootstrapping method was on average close to the true optimal number of variables, but showed considerable spread. Bootstrapping is more accurate in selecting the optimal number of variables than the AIC and BIC alternatives, but this did not translate into a significant difference of the true predictive power. The true predictive power asymptotically converged toward a maximum predictive power for large data sets, and the estimated predictive power converged toward the true predictive power. More than half of the potential predictive power is gained after approximately 200 samples. Our simulations demonstrated severe overfitting (a predicative power lower than that of predicting 50% probability) in a number of small data sets, in particular in data sets with a low number of events (median: 7, 95th percentile: 32). Recognizing overfitting from an inverted sign of the estimated model coefficients has a limited discriminative value.

CONCLUSIONS

Despite considerable spread around the optimal number of selected variables, the bootstrapping method is efficient and accurate for sufficiently large data sets, and guards against overfitting for all simulated cases with the exception of some data sets with a particularly low number of events. An appropriate minimum data set size to obtain a model with high predictive power is approximately 200 patients and more than 32 events. With fewer data samples the true predictive power decreases rapidly, and for larger data set sizes the benefit levels off toward an asymptotic maximum predictive power.

Use of machine learning methods for prediction of acute toxicity in organs at risk following prostate radiotherapy

PURPOSE

The goal of this study is to investigate the advantages of large scale optimization methods vs conventional classification techniques in predicting acute toxicity for urinary bladder and rectum due to prostate irradiation.

METHODS

Clinical and dosimetric data of 321 patients undergoing prostate conformal radiotherapy were recorded. Gastro-intestinal and genito-urinary acute toxicities were scored according to the Radiation Therapy Oncology Group/European Organization for Research and Treatment of Cancer (RTOG/EORTC) scale. Patients were classified in two categories to separate mild (Grade < 2) from severe toxicity levels (Grade > 2). Machine learning methods at different complexity were implemented to predict toxicity as a function of multiple variables. The first approach consisted of a large scale optimization method, based on genetic algorithms (GAs) and artificial neural networks (ANN). The second approach was a binary classifier based on support vector machines (SVM).

RESULTS

The ANN and SVM-based solutions showed comparable prediction accuracy, exhibiting an area under the receiver operating characteristic (ROC) curve of 0.7. Different sensitivity and specificity features were measured for the two approaches. The ANN algorithm showed enhanced sensitivity if combined with appropriate classification criteria.

CONCLUSIONS

The results demonstrate that high sensitivity in toxicity prediction can be achieved with optimized ANNs, that are put forward to represent a valuable support in medical decisions. Future studies will be focused on enlarging the available patient database to increase the reliability of toxicity prediction algorithms and to define optimal classification criteria.

Modeling late rectal toxicities based on a parameterized representation of the 3D dose distribution

Many models exist for predicting toxicities based on dose-volume histograms (DVHs) or dose-surface histograms (DSHs). This approach has several drawbacks as firstly the reduction of the dose distribution to a histogram results in the loss of spatial information and secondly the bins of the histograms are highly correlated with each other. Furthermore, some of the complex nonlinear models proposed in the past lack a direct physical interpretation and the ability to predict probabilities rather than binary outcomes. We propose a parameterized representation of the 3D distribution of the dose to the rectal wall which explicitly includes geometrical information in the form of the eccentricity of the dose distribution as well as its lateral and longitudinal extent. We use a nonlinear kernel-based probabilistic model to predict late rectal toxicity based on the parameterized dose distribution and assessed its predictive power using data from the MRC RT01 trial (ISCTRN 47772397). The endpoints under consideration were rectal bleeding, loose stools, and a global toxicity score. We extract simple rules identifying 3D dose patterns related to a specifically low risk of complication. Normal tissue complication probability (NTCP) models based on parameterized representations of geometrical and volumetric measures resulted in areas under the curve (AUCs) of 0.66, 0.63 and 0.67 for predicting rectal bleeding, loose stools and global toxicity, respectively. In comparison, NTCP models based on standard DVHs performed worse and resulted in AUCs of 0.59 for all three endpoints. In conclusion, we have presented low-dimensional, interpretable and nonlinear NTCP models based on the parameterized representation of the dose to the rectal wall. These models had a higher predictive power than models based on standard DVHs and their low dimensionality allowed for the identification of 3D dose patterns related to a low risk of complication.

Heart irradiation as a risk factor for radiation pneumonitis

PURPOSE

to investigate the potential role of incidental heart irradiation on the risk of radiation pneumonitis (RP) for patients receiving definitive radiation therapy for non-small-cell lung cancer (NSCLC).

MATERIAL AND METHODS

two hundred and nine patient datasets were available for this study. Heart and lung dose-volume parameters were extracted for modeling, based on Monte Carlo-based heterogeneity corrected dose distributions. Clinical variables tested included age, gender, chemotherapy, pre-treatment weight-loss, performance status, and smoking history. The risk of RP was modeled using logistic regression.

RESULTS

the most significant univariate variables were heart related, such as heart heart V65 (percent volume receiving at least 65 Gy) (Spearman Rs = 0.245, p < 0.001). The best-performing logistic regression model included heart D10 (minimum dose to the hottest 10% of the heart), lung D35, and maximum lung dose (Spearman Rs = 0.268, p < 0.0001). When classified by predicted risk, the RP incidence ratio between the most and least risky 1/3 of treatments was 4.8. The improvement in risk modeling using lung and heart variables was better than using lung variables alone.

CONCLUSIONS

these results suggest a previously unsuspected role of heart irradiation in many cases of RP.

Design of and technical challenges involved in a framework for multicentric radiotherapy treatment planning studies

This report introduces a framework for comparing radiotherapy treatment planning in multicentric in silico clinical trials. Quality assurance, data incompatibility, transfer and storage issues, and uniform analysis of results are discussed. The solutions that are given provide a useful guide for the set-up of future multicentric planning studies or public repositories of high quality data.

Comparison of Bayesian network and support vector machine models for two-year survival prediction in lung cancer patients treated with radiotherapy

PURPOSE

Classic statistical and machine learning models such as support vector machines (SVMs) can be used to predict cancer outcome, but often only perform well if all the input variables are known, which is unlikely in the medical domain. Bayesian network (BN) models have a natural ability to reason under uncertainty and might handle missing data better. In this study, the authors hypothesize that a BN model can predict two-year survival in non-small cell lung cancer (NSCLC) patients as accurately as SVM, but will predict survival more accurately when data are missing.

METHODS

A BN and SVM model were trained on 322 inoperable NSCLC patients treated with radiotherapy from Maastricht and validated in three independent data sets of 35, 47, and 33 patients from Ghent, Leuven, and Toronto. Missing variables occurred in the data set with only 37, 28, and 24 patients having a complete data set.

RESULTS

The BN model structure and parameter learning identified gross tumor volume size, performance status, and number of positive lymph nodes on a PET as prognostic factors for two-year survival. When validated in the full validation set of Ghent, Leuven, and Toronto, the BN model had an AUC of 0.77, 0.72, and 0.70, respectively. A SVM model based on the same variables had an overall worse performance (AUC 0.71, 0.68, and 0.69) especially in the Ghent set, which had the highest percentage of missing the important GTV size data. When only patients with complete data sets were considered, the BN and SVM model performed more alike.

CONCLUSIONS

Within the limitations of this study, the hypothesis is supported that BN models are better at handling missing data than SVM models and are therefore more suitable for the medical domain. Future works have to focus on improving the BN performance by including more patients, more variables, and more diversity.