Time-Dependent Deep Learning Prediction of Multiple Sclerosis Disability.

The majority of deep learning models in medical image analysis concentrate on single snapshot timepoint circumstances, such as the identification of current pathology on a given image or volume. This is often in contrast to the diagnostic methodology in radiology where presumed pathologic findings are correlated to prior studies and subsequent changes over time. For multiple sclerosis (MS), the current body of literature describes various forms of lesion segmentation with few studies analyzing disability progression over time. For the purpose of longitudinal time-dependent analysis, we propose a combinatorial analysis of a video vision transformer (ViViT) benchmarked against traditional recurrent neural network of Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) architectures and a hybrid Vision Transformer-LSTM (ViT-LSTM) to predict long-term disability based upon the Extended Disability Severity Score (EDSS). The patient cohort was procured from a two-site institution with 703 patients' multisequence, contrast-enhanced MRIs of the cervical spine between the years 2002 and 2023. Following a competitive performance analysis, a VGG-16-based CNN-LSTM was compared to ViViT with an ablation analysis to determine time-dependency of the models. The VGG16-LSTM predicted trinary classification of EDSS score in 6 years with 0.74 AUC versus the ViViT with 0.84 AUC (p-value < 0.001 per 5 × 2 cross-validation F-test) on an 80:20 hold-out testing split. However, the VGG16-LSTM outperformed ViViT when patients with only 2 years of MRIs (n = 94) (0.75 AUC versus 0.72 AUC, respectively). Exact EDSS classification was investigated for both models using both classification and regression strategies but showed collectively worse performance. Our experimental results demonstrate the ability of time-dependent deep learning models to predict disability in MS using trinary stratification of disability, mimicking clinical practice. Further work includes external validation and subsequent observational clinical trials.

Testicular Dysfunction in Male Childhood Cancer Survivors Treated With Radiation Therapy: A PENTEC Comprehensive Review.

PURPOSE: The male reproductive task force of the Pediatric Normal Tissue Effects in the Clinic (PENTEC) initiative performed a comprehensive review that included a meta-analysis of publications reporting radiation dose-volume effects for risk of impaired fertility and hormonal function after radiation therapy for pediatric malignancies. METHODS AND MATERIALS: The PENTEC task force conducted a comprehensive literature search to identify published data evaluating the effect of testicular radiation dose on reproductive complications in male childhood cancer survivors. Thirty-one studies were analyzed, of which 4 had testicular dose data to generate descriptive scatter plots. Two cohorts were identified. Cohort 1 consisted of pediatric and young adult patients with cancer who received scatter radiation therapy to the testes. Cohort 2 consisted of pediatric and young adult patients with cancer who received direct testicular radiation therapy as part of their cancer therapy. Descriptive scatter plots were used to delineate the relationship between the effect of mean testicular dose on sperm count reduction, testosterone, follicle stimulating hormone (FSH), and luteinizing hormone (LH) levels. RESULTS: Descriptive scatter plots demonstrated a 44% to 80% risk of oligospermia when the mean testicular dose was <1 Gy, but this was recovered by >12 months in 75% to 100% of patients. At doses >1 Gy, the rate of oligospermia increased to >90% at 12 months. Testosterone levels were generally not affected when the mean testicular dose was <0.2 Gy but were abnormal in up to 25% of patients receiving between 0.2 and 12 Gy. Doses between 12 and 19 Gy may be associated with abnormal testosterone in 40% of patients, whereas doses >20 Gy to the testes were associated with a steep increase in abnormal testosterone in at least 68% of patients. FSH levels were unaffected by a mean testicular dose <0.2 Gy, whereas at doses >0.5 Gy, the risk was between 40% and 100%. LH levels were affected at doses >0.5 Gy in 33% to 75% of patients between 10 and 24 months after radiation. Although dose modeling could not be performed in cohort 2, the risk of reproductive toxicities was escalated with doses >10 Gy. CONCLUSIONS: This PENTEC comprehensive review demonstrates important relationships between scatter or direct radiation dose on male reproductive endpoints including semen analysis and levels of FSH, LH, and testosterone.

The Promise and Future of Radiomics for Personalized Radiotherapy Dosing and Adaptation.

Quantitative image analysis, also known as radiomics, aims to analyze large-scale quantitative features extracted from acquired medical images using hand-crafted or machine-engineered feature extraction approaches. Radiomics has great potential for a variety of clinical applications in radiation oncology, an image-rich treatment modality that utilizes computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) for treatment planning, dose calculation, and image guidance. A promising application of radiomics is in predicting treatment outcomes after radiotherapy such as local control and treatment-related toxicity using features extracted from pretreatment and on-treatment images. Based on these individualized predictions of treatment outcomes, radiotherapy dose can be sculpted to meet the specific needs and preferences of each patient. Radiomics can aid in tumor characterization for personalized targeting, especially for identifying high-risk regions within a tumor that cannot be easily discerned based on size or intensity alone. Radiomics-based treatment response prediction can aid in developing personalized fractionation and dose adjustments. In order to make radiomics models more applicable across different institutions with varying scanners and patient populations, further efforts are needed to harmonize and standardize the acquisition protocols by minimizing uncertainties within the imaging data.

Deep learning prediction of post-SBRT liver function changes and NTCP modeling in hepatocellular carcinoma based on DGAE-MRI.

BACKGROUND: Stereotactic body radiation therapy (SBRT) produces excellent local control for patients with hepatocellular carcinoma (HCC). However, the risk of toxicity for normal liver tissue is still a limiting factor. Normal tissue complication probability (NTCP) models have been proposed to estimate the toxicity with the assumption of uniform liver function distribution, which is not optimal. With more accurate regional liver functional imaging available for individual patient, we can improve the estimation and be more patient-specific. PURPOSE: To develop normal tissue complication probability (NTCP) models using pre-/during-treatment (RT) dynamic Gadoxetic Acid-enhanced (DGAE) MRI for adaptation of RT in a patient-specific manner in hepatocellular cancer (HCC) patients who receive SBRT. METHODS: 24 of 146 HCC patients who received SBRT underwent DGAE MRI. Physical doses were converted into EQD2 for analysis. Voxel-by-voxel quantification of the contrast uptake rate (k1) from DGAE-MRI was used to quantify liver function. A logistic dose-response model was used to estimate the fraction of liver functional loss, and NTCP was estimated using the cumulative functional reserve model for changes in Child-Pugh (C-P) scores. Model parameters were calculated using maximum-likelihood estimations. During-RT liver functional maps were predicted from dose distributions and pre-RT k1 maps with a conditional Wasserstein generative adversarial network (cWGAN). Imaging prediction quality was assessed using root-mean-square error (RMSE) and structural similarity (SSIM) metrics. The dose-response and NTCP were fit on both original and cWGAN predicted images and compared using a Wilcoxon signed-rank test. RESULTS: Logistic dose response models for changes in k1 yielded D50 of 35.2 (95% CI: 26.7-47.5) Gy and k of 0.62 (0.49-0.75) for the whole population. The high baseline ALBI (poor liver function) subgroup showed a significantly smaller D50 of 11.7 (CI: 9.06-15.4) Gy and larger k of 0.96 (CI: 0.74-1.22) compared to a low baseline ALBI (good liver function) subgroup of 54.8 (CI: 38.3-79.1) Gy and 0.59 (CI: 0.48-0.74), with p-values of < 0.001 and = 0.008, respectively, which indicates higher radiosensitivity for the worse baseline liver function cohort. Subset analyses were also performed for high/low baseline CP subgroups. The corresponding NTCP models showed good agreement for the fit parameters between cWGAN predicted and the ground-truth during-RT images with no statistical differences for low ALBI subgroup. CONCLUSIONS: NTCP models which incorporate voxel-wise functional information from DGAE-MRI k1 maps were successfully developed and feasibility was demonstrated in a small patient cohort. cWGAN predicted functional maps show promise for estimating localized patient-specific response to RT and warrant further validation in a larger patient cohort.

Predictive Model of Liver Toxicity to Aid the Personalized Selection of Proton Versus Photon Therapy in Hepatocellular Carcinoma.

PURPOSE: Our objective was to develop an externally validated model for predicting liver toxicity after radiation therapy in patients with hepatocellular carcinoma (HCC) that can integrate both photon and proton dose distributions with patient-specific characteristics. METHODS AND MATERIALS: Training data consisted of all patients with HCC treated between 2008 and 2019 at our institution (n = 117, 60%/40% photon/proton). We developed a shallow convolutional neural network (CNN) to predict posttreatment liver dysfunction from the differential dose-volume histogram (DVH) and baseline liver metrics. To reduce bias and improve robustness, we used ensemble learning (CNNE). After a preregistered study analysis plan, we evaluated stability using internal bootstrap resampling and generalizability using a data set from a different institution (n = 88). Finally, we implemented a class activation map method to characterize the critical DVH subregions and benchmarked the model against logistic regression and XGBoost. The models were evaluated using the area under the receiver operating characteristic curve and area under the precision-recall curve. RESULTS: The CNNE model showed similar internal performance and robustness compared with the benchmarks. CNNE exceeded the benchmark models in external validation, with an area under the receiver operating characteristic curve of 0.78 versus 0.55 to 0.70, and an area under the precision-recall curve of 0.6 versus 0.43 to 0.52. The model showed improved predictive power in the photon group, excellent specificity in both modalities, and high sensitivity in the photon high-risk group. Models built solely on DVHs confirm outperformance of the CNNE and indicate that the proposed structure efficiently abstracts features from both proton and photon dose distributions. The activation map method demonstrates the importance of the low-dose bath and its interaction with low liver function at baseline. CONCLUSIONS: We developed and externally validated a patient-specific prediction model for hepatic toxicity based on the entire DVH and clinical factors that can integrate both photon and proton therapy cohorts. This model complements the new American Society for Radiation Oncology clinical practice guidelines and could support value-driven integration of proton therapy into the management of HCC.

Quality Assurance for AI-Based Applications in Radiation Therapy.

Recent advancements in artificial intelligence (AI) in the domain of radiation therapy (RT) and their integration into modern software-based systems raise new challenges to the profession of medical physics experts. These AI algorithms are typically data-driven, may be continuously evolving, and their behavior has a degree of (acceptable) uncertainty due to inherent noise in training data and the substantial number of parameters that are used in the algorithms. These characteristics request adaptive, and new comprehensive quality assurance (QA) approaches to guarantee the individual patient treatment quality during AI algorithm development and subsequent deployment in a clinical RT environment. However, the QA for AI-based systems is an emerging area, which has not been intensively explored and requires interactive collaborations between medical doctors, medical physics experts, and commercial/research AI institutions. This article summarizes the current QA methodologies for AI modules of every subdomain in RT with further focus on persistent shortcomings and upcoming key challenges and perspectives.

Tumor Control Probability After Radiosurgery of Brain Metastases With and Without Retreatment.

PURPOSE: To develop and compare tumor-control probability (TCP) models for single-fraction stereotactic radiosurgery (SRS) for brain metastasis (BMs) with and without retreatment. METHODS AND MATERIALS: We developed three different schemas to model TCP of BMs treated with linear accelerator-based SRS. Dose to 99% of each planning target volume (PTV D99) and 6-month local control were fit using linear-quadratic-linear (LQ-L) models based on equivalent-dose conversions in 2 Gy (EQD2). The M1 schema had separate LQ-L TCP models for initial dose (M1-initial) and retreatment dose (M1-retreat), and the M2 schema had an LQ-L model using the sum of 50% of the initial SRS dose plus the retreatment SRS dose. The M1-initial and M1-retreat schema modeled local control after first SRS to 48 lesions (patients = 22) and second SRS to 46 lesions (patients = 21). The M0 schema included a whole data set of 349 lesions (patients = 136) receiving first SRS (no retreatment and M1-initial). RESULTS: LQ-L models fitted the data well (χ2 = 0.059-0.525 and P = 0.999-1.000). For M0 and M1-retreat, the fitted models EQD250 and γ50 parameters were similar. The LQ-L fitted EQD250 was ∼8.0 Gy for M0 and M1-retreat, ∼24 Gy for M1-initial, and ∼19 Gy for M2. The model fitted γ50 was 0.1 Gy for M0, M1-retreat, and M2 and 0.5 for M1-initial. For the PTV D99 of 10 and 20 Gy, the steepest to shallowest dose-response or largest change in TCP, that is, TCP20Gy - TCP10Gy, was observed in M1-initial (0.49) and M2 (0.17). M0 and M1-retreat showed a similar change in TCP of 0.21. CONCLUSIONS: The model-fitted parameters predicted the recurrent BMs required a higher threshold dose and had a steeper dose-response for first SRS versus second SRS and M0. Alternatively, the recurrent BMs required ∼2 Gy higher predicted PTV D99 dose for first SRS to achieve the same TCP of 0.75 compared with second SRS and M0. Further investigations on larger patient cohorts are needed for validating our findings in predictive modeling of recurrent BMs.

Improved prediction of radiation pneumonitis by combining biological and radiobiological parameters using a data-driven Bayesian network analysis.

Grade 2 and higher radiation pneumonitis (RP2) is a potentially fatal toxicity that limits efficacy of radiation therapy (RT). We wished to identify a combined biomarker signature of circulating miRNAs and cytokines which, along with radiobiological and clinical parameters, may better predict a targetable RP2 pathway. In a prospective clinical trial of response-adapted RT for patients (n = 39) with locally advanced non-small cell lung cancer, we analyzed patients' plasma, collected pre- and during RT, for microRNAs (miRNAs) and cytokines using array and multiplex enzyme linked immunosorbent assay (ELISA), respectively. Interactions between candidate biomarkers, radiobiological, and clinical parameters were analyzed using data-driven Bayesian network (DD-BN) analysis. We identified alterations in specific miRNAs (miR-532, -99b and -495, let-7c, -451 and -139-3p) correlating with lung toxicity. High levels of soluble tumor necrosis factor alpha receptor 1 (sTNFR1) were detected in a majority of lung cancer patients. However, among RP patients, within 2 weeks of RT initiation, we noted a trend of temporary decline in sTNFR1 (a physiological scavenger of TNFα) and ADAM17 (a shedding protease that cleaves both membrane-bound TNFα and TNFR1) levels. Cytokine signature identified activation of inflammatory pathway. Using DD-BN we combined miRNA and cytokine data along with generalized equivalent uniform dose (gEUD) to identify pathways with better accuracy of predicting RP2 as compared to either miRNA or cytokines alone. This signature suggests that activation of the TNFα-NFκB inflammatory pathway plays a key role in RP which could be specifically ameliorated by etanercept rather than current therapy of non-specific leukotoxic corticosteroids.

Quantum deep reinforcement learning for clinical decision support in oncology: application to adaptive radiotherapy.

Subtle differences in a patient's genetics and physiology may alter radiotherapy (RT) treatment responses, motivating the need for a more personalized treatment plan. Accordingly, we have developed a novel quantum deep reinforcement learning (qDRL) framework for clinical decision support that can estimate an individual patient's dose response mid-treatment and recommend an optimal dose adjustment. Our framework considers patients' specific information including biological, physical, genetic, clinical, and dosimetric factors. Recognizing that physicians must make decisions amidst uncertainty in RT treatment outcomes, we employed indeterministic quantum states to represent human decision making in a real-life scenario. We paired quantum decision states with a model-based deep q-learning algorithm to optimize the clinical decision-making process in RT. We trained our proposed qDRL framework on an institutional dataset of 67 stage III non-small cell lung cancer (NSCLC) patients treated on prospective adaptive protocols and independently validated our framework in an external multi-institutional dataset of 174 NSCLC patients. For a comprehensive evaluation, we compared three frameworks: DRL, qDRL trained in a Qiskit quantum computing simulator, and qDRL trained in an IBM quantum computer. Two metrics were considered to evaluate our framework: (1) similarity score, defined as the root mean square error between retrospective clinical decisions and the AI recommendations, and (2) self-evaluation scheme that compares retrospective clinical decisions and AI recommendations based on the improvement in the observed clinical outcomes. Our analysis shows that our framework, which takes into consideration individual patient dose response in its decision-making, can potentially improve clinical RT decision-making by at least about 10% compared to unaided clinical practice. Further validation of our novel quantitative approach in a prospective study will provide a necessary framework for improving the standard of care in personalized RT.

Comparing local control and distant metastasis in NSCLC patients between CyberKnife and conventional SBRT.

BACKGROUND AND PURPOSE: Previous literature suggests that the dose proximally outside the PTV could have an impact on the incidence of distant metastasis (DM) after SBRT in stage I NSCLC patients. We investigated this observation (along with local failure) in deliveries made by different treatment modalities: robotic mounted linac SBRT (CyberKnife) vs conventional SBRT (VMAT/CRT). MATERIALS AND METHODS: This study included 422 stage I NSCLC patients from 2 institutions who received SBRT: 217 treated conventionally and 205 with CyberKnife. The dose behavior outside the PTV of both sub-cohorts were compared by analyzing the mean dose in continuous shells extending 1, 2, 3, …, 100 mm from the PTV. Kaplan-Meier analysis was performed between the two sub-cohorts with respect to DM-free survival and local progression-free survival. A multivariable Cox proportional hazards model was fitted to the combined cohort (n = 422) with respect to DM incidence and local failure. RESULTS: The shell-averaged dose fall-off beyond the PTV was found to be significantly more modest in CyberKnife plans than in conventional SBRT plans. In a 30 mm shell around the PTV, the mean dose delivered with CyberKnife (38.1 Gy) is significantly larger than with VMAT/CRT (22.8 Gy, p<10-8). For 95% of CyberKnife plans, this region receives a mean dose larger than the 21 Gy threshold dose discovered in our previous study. In contrast, this occurs for only 75% of VMAT/CRT plans. The DM-free survival of the entire CyberKnife cohort is superior to that of the 25% of VMAT/CRT patients receiving less than the threshold dose (VMAT/CRT<21Gy), with a hazard ratio of 5.3 (95% CI: 3.0-9.3, p<10-8). The 2 year DM-free survival rates were 87% (95% CI: 81%-91%) and 44% (95% CI: 28%-58%) for CyberKnife and the below-threshold dose conventional cohorts, respectively. A multivariable analysis of the combined cohort resulted in the confirmation that threshold dose was a significant predictor of DM(HR = 0.28, 95% CI: 0.15-0.55, p<10-3) when adjusted for other clinical factors. CyberKnife was also found to be superior to the entire VMAT/CRT with respect to local control (HR = 3.44, CI: 1.6-7.3). The 2-year local progression-free survival rates for the CyberKnife cohort and the VMAT/CRT cohort were 96% (95% CI: 92%-98%) and 88% (95% CI: 82%-92%) respectively. CONCLUSIONS: In standard-of-care CyberKnife treatments, dose distributions that aid distant control are achieved 95% of the time. Although similar doses could be physically achieved by conventional SBRT, this is not always the case with current prescription practices, resulting in worse DM outcomes for 25% of conventional SBRT patients. Furthermore, CyberKnife was found to provide superior local control compared to VMAT/CRT.

Novel Quantitative PET Techniques for Clinical Decision Support in Oncology.

Quantitative image analysis has deep roots in the usage of positron emission tomography (PET) in clinical and research settings to address a wide variety of diseases. It has been extensively employed to assess molecular and physiological biomarkers in vivo in healthy and disease states, in oncology, cardiology, neurology, and psychiatry. Quantitative PET allows relating the time-varying activity concentration in tissues/organs of interest and the basic functional parameters governing the biological processes being studied. Yet, quantitative PET is challenged by a number of degrading physical factors related to the physics of PET imaging, the limitations of the instrumentation used, and the physiological status of the patient. Moreover, there is no consensus on the most reliable and robust image-derived PET metric(s) that can be used with confidence in clinical oncology owing to the discrepancies between the conclusions reported in the literature. There is also increasing interest in the use of artificial intelligence based techniques, particularly machine learning and deep learning techniques in a variety of applications to extract quantitative features (radiomics) from PET including image segmentation and outcome prediction in clinical oncology. These novel techniques are revolutionizing clinical practice and are now offering unique capabilities to the clinical molecular imaging community and biomedical researchers at large. In this report, we summarize recent developments and future tendencies in quantitative PET imaging and present example applications in clinical decision support to illustrate its potential in the context of clinical oncology.

Fluorescence Endomicroscopy Imaging of Mesenchymal Stem Cells in the Rat Lung.

Stem cell therapy has shown great promise for organ repair and regeneration. In the context of lung disease, such as radiation-induced lung damage (RILD) in cancer radiotherapy, mesenchymal stem cells (MSCs) have shown the ability to reduce damage possibly due to their immunomodulatory properties and other unknown mechanisms. However, once MSCs are transplanted into the body, little is known as to their localization or their mechanisms of action. In this work, we proposed, implemented, and validated a fluorescence endomicroscopy (FE) imaging technique that allows for the real-time detection and quantification of transplanted pre-labeled MSCs in vivo and tracking in a rat model. This protocol covers aspects related to MSCs extraction, labeling, FE imaging, and image analysis developed in a RILD rat model but applicable to other biological systems. © 2018 by John Wiley & Sons, Inc.

Can dose outside the PTV influence the risk of distant metastases in stage I lung cancer patients treated with stereotactic body radiotherapy (SBRT)?

BACKGROUND AND PURPOSE: In an era where little is known about the "abscopal" (out-of-the-field) effects of lung SBRT, we investigated correlations between the radiation dose proximally outside the PTV and the risk of cancer recurrence after SBRT in patients with primary stage I non-small cell lung cancer (NSCLC). MATERIALS AND METHODS: This study included 217 stage I NSCLC patients across 2 institutions who received SBRT. Correlations between clinical and dosimetric factors were investigated. The clinical factors considered were distant metastasis (DM), loco-regional control (LRC) and radiation pneumonitis (RP). The dose (converted to EQD2) delivered to regions of varying size directly outside of the PTV was computed. For each feature, area under the curve (AUC) and odds ratios with respect to the outcome parameters DM, LRC and RP were estimated; Kaplan-Meier (KM) analysis was also performed. RESULTS: Thirty-seven (17%) patients developed DM after a median follow-up of 24 months. It was found that the mean dose delivered to a shell-shaped region of thickness 30 mm outside the PTV had an AUC of 0.82. Two years after treatment completion, the rate of DM in patients where the mean dose delivered to this region was higher than 20.8 Gy2 was 5% compared to 60% in those who received a dose lower than 20.8 Gy2. KM analysis resulted in a hazard ratio of 24.2 (95% CI: 10.7, 54.4); p < 10-5. No correlations were found between any factor and either LRC or RP. CONCLUSIONS: The results of this study suggest that the dose received by the region close to the PTV has a significant impact on the risk of distant metastases in stage I NSCLC patients treated with SBRT. If these results are independently confirmed, caution should be taken, particularly when a treatment plan results in a steep dose gradient extending outwards from the PTV.

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.

Image-Guided Fluorescence Endomicroscopy: From Macro- to Micro-Imaging of Radiation-Induced Pulmonary Fibrosis.

Radiation-induced pulmonary fibrosis (RIPF) is a debilitating side effect of radiation therapy (RT) of several cancers including lung and breast cancers. Current clinical methods to assess and monitor RIPF involve diagnostic computed tomography (CT) imaging, which is restricted to anatomical macroscopic changes. Confocal laser endomicroscopy (CLE) or fluorescence endomicroscopy (FE) in combination with a fibrosis-targeted fluorescent probe allows to visualize RIPF in real-time at the microscopic level. However, a major limitation of FE imaging is the lack of anatomical localization of the endomicroscope within the lung. In this work, we proposed and validated the use of x-ray fluoroscopy-guidance in a rat model of RIPF to pinpoint the location of the endomicroscope during FE imaging and map it back to its anatomical location in the corresponding CT image. For varying endomicroscope positions, we observed a positive correlation between CT and FE imaging as indicated by the significant association between increased lung density on CT and the presence of fluorescent fiber structures with FE in RT cases compared to Control. Combining multimodality imaging allows visualization and quantification of molecular processes at specific locations within the injured lung. The proposed image-guided FE method can be extended to other disease models and is amenable to clinical translation for assessing and monitoring fibrotic damage.

Deep reinforcement learning for automated radiation adaptation in lung cancer.

PURPOSE: To investigate deep reinforcement learning (DRL) based on historical treatment plans for developing automated radiation adaptation protocols for nonsmall cell lung cancer (NSCLC) patients that aim to maximize tumor local control at reduced rates of radiation pneumonitis grade 2 (RP2). METHODS: In a retrospective population of 114 NSCLC patients who received radiotherapy, a three-component neural networks framework was developed for deep reinforcement learning (DRL) of dose fractionation adaptation. Large-scale patient characteristics included clinical, genetic, and imaging radiomics features in addition to tumor and lung dosimetric variables. First, a generative adversarial network (GAN) was employed to learn patient population characteristics necessary for DRL training from a relatively limited sample size. Second, a radiotherapy artificial environment (RAE) was reconstructed by a deep neural network (DNN) utilizing both original and synthetic data (by GAN) to estimate the transition probabilities for adaptation of personalized radiotherapy patients' treatment courses. Third, a deep Q-network (DQN) was applied to the RAE for choosing the optimal dose in a response-adapted treatment setting. This multicomponent reinforcement learning approach was benchmarked against real clinical decisions that were applied in an adaptive dose escalation clinical protocol. In which, 34 patients were treated based on avid PET signal in the tumor and constrained by a 17.2% normal tissue complication probability (NTCP) limit for RP2. The uncomplicated cure probability (P+) was used as a baseline reward function in the DRL. RESULTS: Taking our adaptive dose escalation protocol as a blueprint for the proposed DRL (GAN + RAE + DQN) architecture, we obtained an automated dose adaptation estimate for use at ∼2/3 of the way into the radiotherapy treatment course. By letting the DQN component freely control the estimated adaptive dose per fraction (ranging from 1-5 Gy), the DRL automatically favored dose escalation/de-escalation between 1.5 and 3.8 Gy, a range similar to that used in the clinical protocol. The same DQN yielded two patterns of dose escalation for the 34 test patients, but with different reward variants. First, using the baseline P+ reward function, individual adaptive fraction doses of the DQN had similar tendencies to the clinical data with an RMSE = 0.76 Gy; but adaptations suggested by the DQN were generally lower in magnitude (less aggressive). Second, by adjusting the P+ reward function with higher emphasis on mitigating local failure, better matching of doses between the DQN and the clinical protocol was achieved with an RMSE = 0.5 Gy. Moreover, the decisions selected by the DQN seemed to have better concordance with patients eventual outcomes. In comparison, the traditional temporal difference (TD) algorithm for reinforcement learning yielded an RMSE = 3.3 Gy due to numerical instabilities and lack of sufficient learning. CONCLUSION: We demonstrated that automated dose adaptation by DRL is a feasible and a promising approach for achieving similar results to those chosen by clinicians. The process may require customization of the reward function if individual cases were to be considered. However, development of this framework into a fully credible autonomous system for clinical decision support would require further validation on larger multi-institutional datasets.