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

Haralick texture feature analysis for Monte Carlo dose distributions of permanent implant prostate brachytherapy

PURPOSE

Demonstrate quantitative characterization of 3D patient-specific absorbed dose distributions using Haralick texture analysis, and interpret measures in terms of underlying physics and radiation dosimetry.

METHODS

Retrospective analysis is performed for 137 patients who underwent permanent implant prostate brachytherapy using two simulation conditions: "TG186" (realistic tissues including 0-3.8% intraprostatic calcifications; interseed attenuation) and "TG43" (water-model; no interseed attenuation). Five Haralick features (homogeneity, contrast, correlation, local homogeneity, entropy) are calculated using the original Haralick formalism, and a modified approach designed to reduce grey-level quantization sensitivity. Trends in textural features are compared to clinical dosimetric measures (D90; minimum absorbed dose to the hottest 90% of a volume) and changes in patient target volume % intraprostatic calcifications by volume (%IC).

RESULTS

Both original and modified measures quantify the spatial differences in absorbed dose distributions. Strong correlations between differences in textural measures calculated under TG43 and TG186 conditions and %IC are observed for all measures. For example, differences between measures of contrast and correlation increase and decrease respectively as patients with higher levels of %IC are evaluated, reflecting the large differences across adjacent voxels (higher absorbed dose in voxels with calcification) when calculated under TG186 conditions. Conversely, the D90 metric is relatively weakly correlated with textural measures, as it generally does not characterize the spatial distribution of absorbed dose.

CONCLUSION

Patient-specific 3D dose distributions may be quantified using Haralick analysis, and trends may be interpreted in terms of fundamental physics. Promising future directions include investigations of novel treatment modalities and clinical outcomes.

Xerostomia prediction in patients with nasopharyngeal carcinoma during radiotherapy using segmental dose distribution in dosiomics and radiomics models

OBJECTIVES

This study aimed to integrate radiomics and dosiomics features to develop a predictive model for xerostomia (XM) in nasopharyngeal carcinoma after radiotherapy. It explores the influence of distinct feature extraction methods and dose ranges on the performance.

MATERIALS AND METHODS

Data from 363 patients with nasopharyngeal carcinoma were retrospectively analyzed. We pioneered a dose-segmentation strategy, where the overall dose distribution (OD) was divided into four segmental dose distributions (SDs) at intervals of 15 Gy. Features were extracted using manual definition and deep learning, applying OD or SD and integrating radiomics and dosiomics, yielding corresponding feature scores (manually defined radiomics, MDR; manually defined dosiomics, MDD; deep learning-based radiomics, DLR; deep learning-based dosiomics, DLD). Subsequently, 18 models were developed by combining features and model types (random forest and support vector machine).

RESULTS AND CONCLUSION

Under OD, O(DLR_DLD) demonstrated exceptional performance, with an optimal area under the curve (AUC) of 0.81 and an average AUC of 0.71. Within SD, S(DLR_DLD) surpassed the other models, achieving an optimal AUC of 0.90 and an average AUC of 0.85. Therefore, the integration of dosiomics into radiomics can augment predictive efficacy. The dose-segmentation strategy can facilitate the extraction of more profound information. This indicates that ScoreDLR and ScoreMDR were negatively associated with XM, whereas ScoreDLD, derived from SD exceeding 15 Gy, displayed a positive association with XM. For feature extraction, deep learning was superior to manual definition.

Comparison of deep-learning multimodality data fusion strategies in mandibular osteoradionecrosis NTCP modelling using clinical variables and radiation dose distribution volumes

Objective.Normal tissue complication probability (NTCP) modelling is rapidly embracing deep learning (DL) methods, acknowledging the importance of spatial dose information. Finding effective ways to combine information from radiation dose distribution maps (dosiomics) and clinical data involves technical challenges and requires domain knowledge. We propose different multi-modality data fusion strategies to facilitate future DL-based NTCP studies.Approach.Early, joint and late DL multi-modality fusion strategies were compared using clinical and mandibular radiation dose distribution volumes. These were contrasted with single-modality models: a random forest trained on non-image data (clinical, demographic and dose-volume metrics) and a 3D DenseNet-40 trained on image data (mandibular dose distribution maps). The study involved a matched cohort of 92 osteoradionecrosis cases and 92 controls from a single institution.Main results.The late fusion model exhibited superior discrimination and calibration performance, while the join fusion achieved a more balanced distribution of the predicted probabilities. Discrimination performance did not significantly differ between strategies. Late fusion, though less technically complex, lacks crucial inter-modality interactions for NTCP modelling. In contrast, joint fusion, despite its complexity, resulted in a single network training process which included intra- and inter-modality interactions in its model parameter optimisation.Significance.This study is a pioneering effort in comparing different strategies for including image data into DL-based NTCP models in combination with lower dimensional data such as clinical variables. The discrimination performance of such multi-modality NTCP models and the choice of fusion strategy will depend on the distribution and quality of both types of data. Multiple data fusion strategies should be compared and reported in multi-modality NTCP modelling using DL.

Optimising inter-patient image registration for image-based data mining in breast radiotherapy

Background and purpose

Image-based data mining (IBDM) requires spatial normalisation to reference anatomy, which is challenging in breast radiotherapy due to variations in the treatment position, breast shape and volume. We aim to optimise spatial normalisation for breast IBDM.

Materials and methods

Data from 996 patients treated with radiotherapy for early-stage breast cancer, recruited in the REQUITE study, were included. Patients were treated supine (n = 811), with either bilateral or ipsilateral arm(s) raised (551/260, respectively) or in prone position (n = 185). Four deformable image registration (DIR) configurations for extrathoracic spatial normalisation were tested. We selected the best-performing DIR configuration and further investigated two pathways: i) registering prone/supine cohorts independently and ii) registering all patients to a supine reference. The impact of arm positioning in the supine cohort was quantified. DIR accuracy was estimated using Normalised Cross Correlation (NCC), Dice Similarity Coefficient (DSC), mean Distance to Agreement (MDA), 95 % Hausdorff Distance (95 %HD), and inter-patient landmark registration uncertainty (ILRU).

Results

DIR using B-spline and normalised mutual information (NMI) performed the best across all evaluation metrics. Supine-supine registrations yielded highest accuracy (0.98 ± 0.01, 0.91 ± 0.04, 0.23 ± 0.19 cm, 1.17 ± 1.18 cm, 0.51 ± 0.26 cm for NCC, DSC, MDA, 95 %HD, and ILRU), followed by prone-prone and supine-prone registrations. Arm positioning had no significant impact on registration performance. For the best DIR strategy, uncertainty of 0.44 and 0.81 cm in the breast and shoulder regions was found.

Conclusions

B-spline algorithm using NMI and registered supine and prone cohorts independently provides the most optimal spatial normalisation strategy for breast IBDM.

Data Science Opportunities To Improve Radiotherapy Planning and Clinical Decision Making

Radiotherapy aims to achieve a high tumor control probability while minimizing damage to normal tissues. Personalizing radiotherapy treatments for individual patients, therefore, depends on integrating physical treatment planning with predictive models of tumor control and normal tissue complications. Predictive models could be improved using a wide range of rich data sources, including tumor and normal tissue genomics, radiomics, and dosiomics. Deep learning will drive improvements in classifying normal tissue tolerance, predicting intra-treatment tumor changes, tracking accumulated dose distributions, and quantifying the tumor response to radiotherapy based on imaging. Mechanistic patient-specific computer simulations ('digital twins') could also be used to guide adaptive radiotherapy. Overall, we are entering an era where improved modeling methods will allow the use of newly available data sources to better guide radiotherapy treatments.

Dose-toxicity surface histogram-based prediction of radiation dermatitis severity and shape

Objective.This study aimed to develop a new approach to predict radiation dermatitis (RD) by using the skin dose distribution in the actual area of RD occurrence to determine the predictive dose by grade.Approach.Twenty-three patients with head and neck cancer treated with volumetric modulated arc therapy were prospectively and retrospectively enrolled. A framework was developed to segment the RD occurrence area in skin photography by matching the skin surface image obtained using a 3D camera with the skin dose distribution. RD predictive doses were generated using the dose-toxicity surface histogram (DTH) calculated from the skin dose distribution within the segmented RD regions classified by severity. We then evaluated whether the developed DTH-based framework could visually predict RD grades and their occurrence areas and shapes according to severity.Main results.The developed framework successfully generated the DTH for three different RD severities: faint erythema (grade 1), dry desquamation (grade 2), and moist desquamation (grade 3); 48 DTHs were obtained from 23 patients: 23, 22, and 3 DTHs for grades 1, 2, and 3, respectively. The RD predictive doses determined using DTHs were 28.9 Gy, 38.1 Gy, and 54.3 Gy for grades 1, 2, and 3, respectively. The estimated RD occurrence area visualized by the DTH-based RD predictive dose showed acceptable agreement for all grades compared with the actual RD region in the patient. The predicted RD grade was accurate, except in two patients.Significance. The developed DTH-based framework can classify and determine RD predictive doses according to severity and visually predict the occurrence area and shape of different RD severities. The proposed approach can be used to predict the severity and shape of potential RD in patients and thus aid physicians in decision making.

Predictive Factors for Toxicity After Primary Chemoradiation for Locally Advanced Cervical Cancer: A Systematic Review

PURPOSE

Women with locally advanced cervical cancer (LACC) undergoing primary platinum-based chemoradiotherapy and brachytherapy often experience toxicities. Normal-tissue complication probability (NTCP) models quantify toxicity risk and aid in optimizing radiation therapy to minimize side effects. However, it is unclear which predictors to include in an NTCP model. The aim of this systematic review was to provide an overview of the identified predictors contributing to gastrointestinal (GI), genitourinary (GU), and vaginal toxicities and insufficiency fractures for LACC.

METHODS AND MATERIALS

A systematic search was performed and articles evaluating the relationship between predictors and toxicities in women with LACC treated with primary chemoradiation were included. The Quality In Prognosis Studies tool was used to assess risk of bias, with high-risk studies being excluded from further analysis. Relationships between dose-volume parameters, patient and treatment characteristics, and toxicity endpoints were analyzed.

RESULTS

Seventy-three studies were identified. Twenty-six had a low or moderate risk of bias and were therefore included. Brachytherapy-related dose-volume parameters of the GI tract, including rectum and bowel equivalent dose in 2 Gy fractions (EQD2) D2 cm3, were frequently related to toxicities, unlike GU dose-volume parameters. Furthermore, (recto)vaginal point doses predicted toxicities. Few studies evaluated external beam radiation therapy dose-volume parameters and identified rectum EQD2 V30 Gy, V40 Gy, and V55 Gy, bowel and bladder EQD2 V40 Gy as toxicity predictors. Also, total reference air kerma and vaginal reference length were associated with toxicities. Relationships between patient characteristics and GI toxicity were inconsistent. The extent of vaginal involvement at diagnosis, baseline symptoms, and obesity predicted GU or vaginal toxicities. Only 1 study evaluated insufficiency fractures and demonstrated lower pretreatment bone densities to be associated.

CONCLUSIONS

This review detected multiple candidate predictors of toxicity. Larger studies should consider insufficiency fractures, assess dose levels from external beam radiation therapy, and quantify the relationship between the predictors and treatment-related toxicities in women with LACC to further facilitate NTCP model development for clinical use.

Voxel-based analysis: Roadmap for clinical translation

Voxel-based analysis (VBA) allows the full, 3-dimensional, dose distribution to be considered in radiotherapy outcome analysis. This provides new insights into anatomical variability of pathophysiology and radiosensitivity by removing the need for a priori definition of organs assumed to drive the dose response associated with patient outcomes. This approach may offer powerful biological insights demonstrating the heterogeneity of the radiobiology across tissues and potential associations of the radiotherapy dose with further factors. As this methodological approach becomes established, consideration needs to be given to translating VBA results to clinical implementation for patient benefit. Here, we present a comprehensive roadmap for VBA clinical translation. Technical validation needs to demonstrate robustness to methodology, where clinical validation must show generalisability to external datasets and link to a plausible pathophysiological hypothesis. Finally, clinical utility requires demonstration of potential benefit for patients in order for successful translation to be feasible. For each step on the roadmap, key considerations are discussed and recommendations provided for best practice.

Selective sparing of bladder and rectum sub-regions in radiotherapy of prostate cancer combining knowledge-based automatic planning and multicriteria optimization

Background and Purpose

The association between dose to selected bladder and rectum symptom-related sub-regions (SRS) and late toxicity after prostate cancer radiotherapy has been evidenced by voxel-wise analyses. The aim of the current study was to explore the feasibility of combining knowledge-based (KB) and multi-criteria optimization (MCO) to spare SRSs without compromising planning target volume (PTV) dose delivery, including pelvic-node irradiation.

Materials and Methods

Forty-five previously treated patients (74.2 Gy/28fr) were selected and SRSs (in the bladder, associated with late dysuria/hematuria/retention; in the rectum, associated with bleeding) were generated using deformable registration. A KB model was used to obtain clinically suitable plans (KB-plan). KB-plans were further optimized using MCO, aiming to reduce dose to the SRSs while safeguarding target dose coverage, homogeneity and avoiding worsening dose volume histograms of the whole bladder, rectum and other organs at risk. The resulting MCO-generated plans were examined to identify the best-compromise plan (KB + MCO-plan).

Results

The mean SRS dose decreased in almost all patients for each SRS. D1% also decreased in the large majority, less frequently for dysuria/bleeding SRS. Mean differences were statistically significant (p < 0.05) and ranged between 1.3 and 2.2 Gy with maximum reduction of mean dose up to 3-5 Gy for the four SRSs. The better sparing of SRSs was obtained without compromising PTVs coverage.

Conclusions

Selectively sparing SRSs without compromising PTV coverage is feasible and has the potential to reduce toxicities in prostate cancer radiotherapy. Further investigation to better quantify the expected risk reduction of late toxicities is warranted.

Voxel-wise analysis: A powerful tool to predict radio-induced toxicity and potentially perform personalised planning in radiotherapy

Dose - volume histograms have been historically used to study the relationship between the planned radiation dose and healthy tissue damage. However, this approach considers neither spatial information nor heterogenous radiosensitivity within organs at risk, depending on the tissue. Recently, voxel-wise analyses have emerged in the literature as powerful tools to fully exploit three-dimensional information from the planned dose distribution. They allow to identify anatomical subregions of one or several organs in which the irradiation dose is associated with a given toxicity. These methods rely on an accurate anatomical alignment, usually obtained by means of a non-rigid registration. Once the different anatomies are spatially normalised, correlations between the three-dimensional dose and a given toxicity can be explored voxel-wise. Parametric or non-parametric statistical tests can be performed on every voxel to identify the voxels in which the dose is significantly different between patients presenting or not toxicity. Several anatomical subregions associated with genitourinary, gastrointestinal, cardiac, pulmonary or haematological toxicity have already been identified in the literature for prostate, head and neck or thorax irradiation. Voxel-wise analysis appears therefore first particularly interesting to increase toxicity prediction capability by identifying specific subregions in the organs at risk whose irradiation is highly predictive of specific toxicity. The second interest is potentially to decrease the radio-induced toxicity by limiting the dose in the predictive subregions, while not decreasing the dose in the target volume. Limitations of the approach have been pointed out.

Haralick texture feature analysis for characterization of specific energy and absorbed dose distributions across cellular to patient length scales

Objective.To investigate an approach for quantitative characterization of the spatial distribution of dosimetric data by introducing Haralick texture feature analysis in this context.Approach.Monte Carlo simulations are used to generate 3D arrays of dosimetric data for 2 scenarios: (1) cell-scale microdosimetry: specific energy (energy imparted per unit mass) in cell-scale targets irradiated by photon spectra (¹²⁵I,¹⁹²Ir, 6 MV); (2) tumour-scale dosimetry: absorbed dose in voxels for idealized models of¹²⁵I permanent implant prostate brachytherapy, considering 'TG186' (realistic tissues including 0% to 5% intraprostatic calcifications; interseed attenuation) and 'TG43' (water model, no interseed attenuation) conditions. Five prominent Haralick features (homogeneity, contrast, correlation, local homogeneity, entropy) are computed and trends are interpreted using fundamental radiation physics.Main results.In the cell-scale scenario, the Haralick measures quantify differences in 3D specific energy distributions due to source spectra. For example, contrast and entropy are highest for¹²⁵I reflecting the large variations in specific energy in adjacent voxels (photoelectric interactions; relatively short range of electrons), while 6 MV has the highest homogeneity with smaller variations in specific energy between voxels (Compton scattering dominates; longer range of electrons). For the tumour-scale scenario, the Haralick measures quantify differences due to TG186/TG43 simulation conditions and the presence of calcifications. For example, as calcifications increase from 0% to 5%, contrast increases while correlation decreases, reflecting the large differences in absorbed dose in adjacent voxels (higher absorbed dose in voxels with calcification due to photoelectric interactions).Significance.Haralick texture analysis provides a quantitative method for the characterization of 3D dosimetric distributions across cellular to tumour length scales, with promising future applications including analyses of multiscale tissue models, patient-specific data, and comparison of treatment approaches.

A Dosiomics Analysis Based on Linear Energy Transfer and Biological Dose Maps to Predict Local Recurrence in Sacral Chordomas after Carbon-Ion Radiotherapy

Carbon Ion Radiotherapy (CIRT) is one of the most promising therapeutic options to reduce Local Recurrence (LR) in Sacral Chordomas (SC). The aim of this work is to compare the performances of survival models fed with dosiomics features and conventional DVH metrics extracted from relative biological effectiveness (RBE)-weighted dose (D_RBE) and dose-averaged Linear Energy Transfer (LET_d) maps, towards the identification of possible prognostic factors for LR in SC patients treated with CIRT. This retrospective study included 50 patients affected by SC with a focus on patients that presented a relapse in a high-dose region. Survival models were built to predict both LR and High-Dose Local Recurrencies (HD-LR). The models were evaluated through Harrell Concordance Index (C-index) and patients were stratified into high/low-risk groups. Local Recurrence-free Kaplan-Meier curves were estimated and evaluated through log-rank tests. The model with highest performance (median(interquartile-range) C-index of 0.86 (0.22)) was built on features extracted from LET_d maps, with D_RBE models showing promising but weaker results (C-index of 0.83 (0.21), 0.80 (0.21)). Although the study should be extended to a wider patient population, LET_d maps show potential as a prognostic factor for SC HD-LR in CIRT, and dosiomics appears to be the most promising approach against more conventional methods (e.g., DVH-based).

Technical note: rtdsm-An open-source software for radiotherapy dose-surface map generation and analysis

BACKGROUND

Dose-outcome studies in radiation oncology have historically excluded spatial information due to dose-volume histograms being the most dominant source of dosimetric information. In recent years, dose-surface maps (DSMs) have become increasingly popular for characterization of spatial dose distributions and identification of radiosensitive subregions for hollow organs. However, methodological variations and lack of open-source, publicly offered code-sharing between research groups have limited reproducibility and wider adoption.

PURPOSE

This paper presents rtdsm, an open-source software for DSM calculation with the intent to improve the reproducibility of and the access to DSM-based research in medical physics and radiation oncology.

METHODS

A literature review was conducted to identify essential functionalities and prevailing calculation approaches to guide development. The described software has been designed to calculate DSMs from DICOM data with a high degree of user customizability and to facilitate DSM feature analysis. Core functionalities include DSM calculation, equivalent dose conversions, common DSM feature extraction, and simple DSM accumulation.

RESULTS

A number of use cases were used to qualitatively and quantitatively demonstrate the use and usefulness of rtdsm. Specifically, two DSM slicing methods, planar and noncoplanar, were implemented and tested, and the effects of method choice on output DSMs were demonstrated. An example comparison of DSMs from two different treatments was used to highlight the use cases of various built-in analysis functions for equivalent dose conversion and DSM feature extraction.

CONCLUSIONS

We developed and implemented rtdsm as a standalone software that provides all essential functionalities required to perform a DSM-based study. It has been made freely accessible under an open-source license on Github to encourage collaboration and community use.

Extending the relative seriality formalism for interpretable deep learning of normal tissue complication probability models

We formally demonstrate that the relative seriality (RS) model of normal tissue complication probability (NTCP) can be recast as a simple neural network with one convolutional and one pooling layer. This approach enables us to systematically construct deep relative seriality networks (DRSNs), a new class of mechanistic generalizations of the RS model with radiobiologically interpretable parameters amenable to deep learning. To demonstrate the utility of this formulation, we analyze a simplified example of xerostomia due to irradiation of the parotid gland during alpha radiopharmaceutical therapy. Using a combination of analytical calculations and numerical simulations, we show for both the RS and DRSN cases that the ability of the neural network to generalize without overfitting is tied to ‘stiff’ and ‘sloppy’ directions in the parameter space of the mechanistic model. These results serve as proof-of-concept for radiobiologically interpretable deep learning of NTCP, while simultaneously yielding insight into how such techniques can robustly generalize beyond the training set despite uncertainty in individual parameters.

Cluster model incorporating heterogeneous dose distribution of partial parotid irradiation for radiotherapy induced xerostomia prediction with machine learning methods

PURPOSE

A cluster model incorporating heterogeneous dose distribution within the parotid gland was developed and validated retrospectively for radiotherapy (RT) induced xerostomia prediction with machine learning (ML) techniques.

METHODS

Sixty clusters were obtained at 1 Gy step size with threshold doses ranging from 1 to 60 Gy, for each of the enrolled 155 patients with HNC from three institutions. Feature clusters were selected with the neighborhood component analysis (NCA) and subsequently fed into four supervised ML models for xerostomia prediction comparison: support vector machines (SVM), k-nearest neighbor (kNN), naïve Bayes (NB), and random forest (RF). The predictive performance of each model was evaluated using cross validation resampling with the area-under-the-curves (AUC) of the receiver-operating-characteristic (ROC). The xerostomia predicting capacity using testing data was assessed with accuracy, sensitivity, and specificity for these models and three cluster connectivity choices. Mean dose based logistic regression served as the benchmark for evaluation.

RESULTS

Feature clusters identified by NCA fell in three threshold dose ranges: 5-15Gy, 25-35Gy, and 45-50Gy. Mean dose predictive power was 15% lower than that of the cluster model using the logistic regression classifier. Model validation demonstrated that kNN model outperformed slightly other three models but no substantial difference was observed. Applying the fine-tuned models to testing data yielded that the mean accuracy from SVM, kNN and NB models were between 0.68 and 0.7 while that of RF was ∼0.6. SVM model yielded the best sensitivity (0.76) and kNN model delivered consistent sensitivity and specificity. This is consistent with cross validation. Clusters calculated with three connectivity choices exhibited minimally different predictions.

CONCLUSION

Compared to mean dose, the proposed cluster model has shown its improvement as the xerostomia predictor. When combining with ML techniques, it could provide a clinically useful tool for xerostomia prediction and facilitate decision making during radiotherapy planning for patients with HNC.

Can dose convolution modelling explain bath and shower effects in rat spinal cord?

Objective. ‘Bath and shower’ effects were first seen in proton irradiations of rat spinal cord, where a low dose ‘bath’ reduced the smaller field ‘shower’ dose needed for limb paralysis giving the appearance of sensitisation of the cord or disproportionate response. This was difficult to reconcile with existing tissue complication models. The purpose of this investigation is to explore a different approach using a dose convolution algorithm to model the 50% isoeffect endpoint. Approach. Bath and shower dose distributions were convolved with Gaussian functions with widths specified by the σ parameter. The hypothesis was that the maximum value from the convolved distributions was constant for isoeffect across the modelled scenarios. A simpler field length dependent relative biological effectiveness (FLRBE) approach was also used for a subset of the data which gave results independent of σ. Main results. The maximum values from the convolved distributions were constant within ±17% across the bath and shower experiments for σ = 3.5 mm, whereas the maximum dose varied by a factor of four. The FLRBE results were also within ±14% confirming the validity of the dose convolution approach. Significance. A simple approach using dose convolution modelling of the 50% isotoxicity gave compelling consistency with the full range of bath and shower results, while the FLRBE approach confirmed the results for the symmetric field data. Convolution modelling and the effect of time interval were consistent with a signalling factor diffusion mechanism such as the ‘bystander effect’. The results suggest biological effectiveness is reduced for very small field sizes, requiring a higher isoeffect dose. By implication, the bath dose does not sensitise the cord to the shower dose; when biological effectiveness is accounted for, a small increase in the bath dose requires a significantly larger reduction in shower dose for isoeffect.

Radiation-Induced Esophagitis in Non-Small-Cell Lung Cancer Patients: Voxel-Based Analysis and NTCP Modeling

Simple Summary

Radiation-induced esophagitis (RE) is a common dose-limiting complication associated with concurrent chemoradiation therapy for Non-Small-Cell Lung Cancer (NSCLC), and a wide range of esophageal dosimetric parameters have been described as predictive of RE. In this study, we characterize the risk of RE for NSCLC patients enrolled in a prospective trial comparing intensity-modulated RT versus passive scattering proton therapy for locally advanced NSCLC. Dose patterns associated with RE were analyzed by applying voxel-based analysis approaches, and predictive models for RE were finally investigated. Two predictive models for acute RE with good cross-validated predictive performances and discrimination capability were developed (thoracic esophageal model: ROC-AUC = 0.73; whole esophagus model: ROC-AUC = 0.70).

Abstract

The aim of our study is to characterize the risk of radiation-induced esophagitis (RE) in a cohort of Non-Small-Cell Lung Cancer (NSCLC) patients treated with concurrent chemotherapy and photon/proton therapy. For each patient, the RE was graded according to the CTCAE v.3. The esophageal dose-volume histograms (DVHs) were extracted. Voxel-based analyses (VBAs) were performed to assess the spatial patterns of the dose differences between patients with and without RE of grade ≥ 2. Two hierarchical NTCP models were developed by multivariable stepwise logistic regression based on non-dosimetric factors and on the DVH metrics for the whole esophagus and its anatomical subsites identified by the VBA. In the 173 analyzed patients, 76 (44%) developed RE of grade ≥ 2 at a median follow-up time of 31 days. The VBA identified regions of significant association between dose and RE in a region encompassing the thoracic esophagus. We developed two NTCP models, including the RT modality and a dosimetric factor: V_55Gy for the model related to the whole esophagus, and the mean dose for the model designed on the thoracic esophagus. The cross-validated performance showed good predictions for both models (ROC-AUC of 0.70 and 0.73, respectively). The only slight improvement provided by the analysis of the thoracic esophageal subsites might be due to the relevant sparing of cervical and lower thoracic esophagus in the analyzed cohort. Further studies on larger cohorts and a more heterogeneous set of dose distributions are needed to validate these preliminary findings and shed further light on the spatial patterns of RE development.

Deep Learning for Radiotherapy Outcome Prediction Using Dose Data - A Review

Artificial intelligence, and in particular deep learning using convolutional neural networks, has been used extensively for image classification and segmentation, including on medical images for diagnosis and prognosis prediction. Use in radiotherapy prognostic modelling is still limited, however, especially as applied to toxicity and tumour response prediction from radiation dose distributions. We review and summarise studies that applied deep learning to radiotherapy dose data, in particular studies that utilised full three-dimensional dose distributions. Ten papers have reported on deep learning models for outcome prediction utilising spatial dose information, whereas four studies used reduced dimensionality (dose volume histogram) information for prediction. Many of these studies suffer from the same issues that plagued early normal tissue complication probability modelling, including small, single-institutional patient cohorts, lack of external validation, poor data and model reporting, use of late toxicity data without taking time-to-event into account, and nearly exclusive focus on clinician-reported complications. They demonstrate, however, how radiation dose, imaging and clinical data may be technically integrated in convolutional neural networks-based models; and some studies explore how deep learning may help better understand spatial variation in radiosensitivity. In general, there are a number of issues specific to the intersection of radiotherapy outcome modelling and deep learning, for example translation of model developments into treatment plan optimisation, which will require further combined effort from the radiation oncology and artificial intelligence communities.