Vargas, Carlos

Abstract Background Accurate and efficient dose calculation is essential for online adaptive planning in proton therapy. Deep learning (DL) has shown promising dose prediction results for pencil beam scanning proton therapy (PBSPT) in recent years, but existing DL‐based dose prediction methods still suffer from limited generalizability and an inability to effectively handle outlier clinical cases. This may lead to inaccurate dose delivery to targets or excessive irradiation to organs at risk (OARs), thereby compromising the safety and efficacy of online adaptive proton therapy. Purpose To design a physics‐aware and generalizable AI‐based PBSPT dose prediction method that incorporates underlying physics to enhance generalizability, particularly in handling outlier clinical cases. Methods This study analyzed PBSPT plans of 103 prostate (93 for training and 10 for testing) and 78 lung cancer patients (68 for training and 10 for testing) from our institution, with each case comprising CT images and structure sets. Using the doses generated by our Monte Carlo‐based dose engine as the reference standard, we compared three methods: the region of interest (ROI)‐based method, the beam mask and sliding window method, and the proposed noisy probing dose method, which rapidly generates a low‐statistics dose via uniformly weighted spots on an expanded spot‐placement target volume without optimization. To evaluate the generalizability of these methods to rare treatment planning scenarios, 12 cases with uncommon beam angles or prescription doses were used to assess their performance, which was evaluated using dose‐volume histogram (DVH) indices, 3D Gamma passing rates (3%/2 mm/10%), and Dice coefficients for dose agreement, while prediction times were measured to gauge model efficiency. Results The proposed noisy probing dose method consistently outperformed the ROI‐based and beam mask baselines across all evaluated metrics, with more accurate dose agreement and superior generalizability. For DVH indices, the noisy probing dose method achieved the smallest deviation in clinical target volume (CTV) dose coverage: in prostate cancer, CTV D98 deviation was reduced by 45% (from 0.53 ± 0.22 Gy [RBE] for ROI‐based) and 29% (from 0.41 ± 0.28 Gy [RBE] for beam mask) to 0.29 ± 0.06 Gy [RBE]; in lung cancer, similar improvements were observed, with CTV D98 deviations reduced to 0.34 ± 0.12 Gy [RBE]. The 3D Gamma passing rates improved to 99.65% ± 1.15% for prostate targets and 97.04% ± 1.17% for lung targets. The dice coefficients of the 90% iso‐dose lines were also the highest with the noisy probing dose method (prostate: 0.983 ± 0.005; lung: 0.967 ± 0.01). For the 12 outlier cases, the noisy probing dose method maintained superior generalizability, yielding higher 3D Gamma passing rates (prostate targets: 96.79% ± 0.83%, OARs: 94.29% ± 1.01%; lung targets: 93.38% ± 1.34%, OARs: 93.95% ± 1.32%), demonstrating robust generalizability to rare clinical scenarios. The dose predictions for all testing cases were completed within 0.3 seconds. Conclusions A novel noisy probing dose method was proposed for PBSPT dose prediction in prostate and lung cancer patients. By embedding more proton‐specific physics, this method demonstrated an improvement in the generalizability of dose prediction.

Abstract Background The radiation oncology clinical practice involves many steps relying on the dynamic interplay of abundant text data. Large language models have displayed remarkable capabilities in processing complex text information. But their direct applications in specific fields like radiation oncology remain underexplored. Purpose This study aims to investigate whether fine‐tuning LLMs with domain knowledge can improve the performance on Task (1) treatment regimen generation, Task (2) treatment modality selection (photon, proton, electron, or brachytherapy), and Task (3) ICD‐10 code prediction in radiation oncology. Methods Data for 15 724 patient cases were extracted. Cases where patients had a single diagnostic record, and a clearly identifiable primary treatment plan were selected for preprocessing and manual annotation to have 7903 cases of the patient diagnosis, treatment plan, treatment modality, and ICD‐10 code. Each case was used to construct a pair consisting of patient diagnostics details and an answer (treatment regimen, treatment modality, or ICD‐10 code, respectively) for the supervised fine‐tuning of these three tasks. Open source LLaMA2‐7B and Mistral‐7B models were utilized for the fine‐tuning with the Low‐Rank Approximations method. Accuracy and ROUGE‐1 score were reported for the fine‐tuned models and original models. Clinical evaluation was performed on Task (1) by radiation oncologists, while precision, recall, and F‐1 score were evaluated for Task (2) and (3). One‐sided Wilcoxon signed‐rank tests were used to statistically analyze the results. Results Fine‐tuned LLMs outperformed original LLMs across all tasks with p value ≤ 0.001. Clinical evaluation demonstrated that over 60% of the fine‐tuned LLMs‐generated treatment regimens were clinically acceptable. Precision, recall, and F1‐score showed improved performance of fine‐tuned LLMs. Conclusion Fine‐tuned LLMs demonstrated statistically significant improvements over original LLMs upon three clinically important tasks in radiation oncology. This study explored the feasibility of applying fine‐tuned LLMs in radiation oncology, inspiring further development of utilizing LLMs to assist with radiation oncology tasks.

Abstract Introduction: Radiotherapy (RT)-induced acute esophagitis (AE) is a common side effect in lung cancer patients receiving RT, which can significantly impact their quality of life. This highlights the need for a predictive model that can estimate AE risk in advance using pretreatment imaging data. However, collecting sufficient data for model development is often resource-intensive and costly. Additionally, acquiring a homogeneous training dataset (e.g., from a single center or modality) is not always feasible. This study aimed to develop an automated artificial intelligence-based 3D vision foundation model (VFM) combining standard-of-care planning computed tomography (pCT) and planned radiation dose maps to predict grade II or higher AE. Materials and methods: This study included 246 patients with non-small cell lung cancer who underwent either image-guided radiation treatment (IMRT) or proton-beam radiation therapy. The endpoint was grade two or higher AE (33% positive and 67% negative). The IMRT group consisted of 182 patients from a single center, whereas the proton therapy group included 64 patients from the Proton Collaborative Group trial from 11 different institutions. For each patient, pCT, dose maps, and radiotherapy segmentations were available. The VFM was created using a transformer pretrained using a large number of unlabeled volumetric 3D CTs from patients with varied diseases through a self-supervised learning approach, which extracts useful features directly from images. Next, the pretrained encoder was combined with fully connected classification layers and fine-tuned using stratified 5-fold cross-validation with 20 patients set aside for testing. We developed two models: a CT only and CT + dose model. Model performance was assessed using the area under the curve (AUC), specificity, and sensitivity. Results: In cross-validation, adding dose information to CT (CT+Dose) improved model performance, increasing AUC from 0.72 ± 0.08 to 0.76 ± 0.10 and specificity from 0.82 ± 0.09 to 0.87 ± 0.04, while sensitivity remained the same. On the independent test set, CT+Dose showed a marked improvement over CT only (AUC: 0.82 vs. 0.59; sensitivity: 0.57 vs. 0.14), with a slight decrease in specificity (0.77 vs. 1.00). Conclusion: A VFM model combining CT and radiation dose showed the capability to predict AE more accurately with higher specificity. Further studies on larger cohorts of testing patients are planned to assess model generalization. Citation Format: Chloe Min Seo Choi, Jue Jiang, Nikhil Mankuzhy, John Chang, Jing Zeng, Carlos Vargas, James Urbanic, Isabella Choi, Mark Mcdonald, James Gray, Joseph Deasy, Maria Thor, Charles Simone, Harini Veeraraghavan. Prediction of radiotherapy-Induced esophagitis in non-small cell lung cancer using a 3D vision foundation model [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B022.

Abstract Background Inter‐fractional and intra‐fractional positioning and motion monitoring based on implanted fiducials through kilovoltage (kV) imaging is a cost‐effective approach to enhance intact prostate radiotherapy treatment accuracy. However, comprehensive studies for three‐dimensional (3D) position extraction and the corrective action threshold are lacking. Purpose To develop and verify a fiducial motion monitoring system with 3D position information based on triggered kV images taken during treatment delivery, and to comprehensively evaluate the recommended thresholds for corrective action during treatment. Methods An in‐house fiducial triangulation algorithm was developed to monitor fiducial positions using subsequent kV triggered images. The precision of fiducial triangulation and motion detection was validated on a pelvis phantom with four fiducial inserts. A retrospective study was conducted on prostate cancer patients who received either moderately hypofractionated radiotherapy (MHRT, n = 16) or stereotactic body radiotherapy (SBRT, n = 12), categorized by endo‐rectal balloon (ERB) use. Intra‐fractional positions relative to isocenter were computed, and these positions were analyzed in comparison to the original marker locations of the patient's treatment plan. A linear regression fit was used per fraction to determine correlation coefficients between motions in the left–right (LR), superior–inferior (SI), and anterior–posterior (AP) directions. Average fraction time (AFT) of treatment was reported based on the observed average time spent on steps in the workflow and the percentage of fiducials caught outside of the current two‐dimensional (2D) tolerance threshold of 5 mm. The observed 3D offsets were used to estimate AFT for various 3D tolerance thresholds. Results Overall, our triangulation method proved to be very precise for static cases, where phantom measurements revealed a fiducial position triangulation precision of mm for stationary targets, but had a spread of mm for targets with 1 mm of motion. Employing triangulation, true motion exceeding 5 mm was detected above 4.0 mm in magnitude 90.9% of the time, a noticeable improvement in comparison to the 28%–65% successful detection rate reported by 2D projection methods. Most detection errors were attributed to depth disparity. Similar to other reports, correlation coefficients for intra‐fractional motion generally indicated no LR/SI, no LR/AP, and weak positive SI/AP correlations for MHRT and SBRT patients. Offsets beyond a 3D tolerance threshold of 5 mm were observed at a rate of 4.25%–5.25%, while the 5 mm 2D out‐of‐tolerance catch‐rate was 1.6%. The AFT was 8.1–8.2 min using the 2D tolerance threshold of 5 mm. In comparison, the estimated AFT for the proposed 3D tolerance monitoring of offsets beyond 2–5 mm was slightly higher at 8.2–10.9 min due to the higher amount of out‐of‐tolerance instances for the higher precision intra‐fractional motion management. Conclusions Our study showcases a promising subsequent kV‐based triangulation method for intra‐fractional prostate motion monitoring. Acquiring 3D motion information results in higher out‐of‐tolerance catch‐rates particularly in the depth dimension of the kV images, which is perpendicular to the treatment beam. Failure to properly observe and catch these offsets would result in sub‐optimal conformality and accuracy of the dose delivery.

Abstract BACKGROUND Standard of care treatment for recurrent malignant glioma after prior radiotherapy (RT) is not well established. Proton therapy (PT) is increasingly used for reirradiation (ReRT) at the time of recurrence; however, treatment outcomes, toxicity, and prognostic factors for patients treated with PT-ReRT remain poorly defined. METHODS The prospective, multi-institutional Proton Collaborative Group (PCG) registry was queried for patients with malignant glioma who previously received RT and underwent PT-ReRT between 7/2011-12/2023. Overall survival (OS) and progression-free survival (PFS) were assessed using the Kaplan-Meier method, and Cox proportional hazards regression was used for uni- and multivariate analysis (UVA and MVA). Dose was calculated using equivalent dose in 2-Gy fractions (EQD2). RESULTS 143 consecutive patients were identified. Median follow-up was 11.2 months and median time interval (TI) from prior RT (median 58.5 Gy) to PT-ReRT (median 44.6 Gy) was 42.4 months. Primary histology was glioblastoma (n=74), astrocytoma (n=35), and oligodendroglioma (n=34). Median PFS and OS were 8.1 and 11.2 months, respectively. On UVA, improved OS was associated with oligodendroglioma (HR 0.37) and astrocytoma histology (HR 0.53) compared to glioblastoma, TI >60 months (HR 0.2), CNS WHO Grade 2 compared to 4 (HR 0.42), and ECOG performance status 0 (HR 0.47). On MVA, improved OS remained associated only with oligodendroglioma (HR 0.42) and TI >60 months (HR 0.49). Acute and Late Grade 3 toxicity occurred in 7% and 4% of patients. Acute Grade 3 toxicity was associated with poor performance status (HR 16.15). CONCLUSIONS In the largest series of glioma PT-ReRT reported to date, retreatment appears well tolerated with variable outcomes based on clinical prognostic factors. Toxicity rates were acceptable compared to historical photon-based literature despite a high median prescription dose.

AbstractPurposeTo assess the practicality of employing a commercial knowledge‐based planning tool (RapidPlan) to generate adapted intact prostate and prostate bed volumetric modulated arc therapy (VMAT) plans on iterative cone‐beam computed tomography (iCBCT) datasets.Methods and materialsIntact prostate and prostate bed RapidPlan models were trained utilizing planning data from 50 and 44 clinical cases, respectively. To ensure that refined models were capable of producing adequate clinical plans with a single optimization, models were tested with 50 clinical planning CT datasets by comparing dose‐volume histogram (DVH) and plan quality metric (PQM) values between clinical and RapidPlan‐generated plans. The RapidPlan tool was then used to retrospectively generate adapted VMAT plans on daily iCBCT images for 20 intact prostate and 15 prostate bed cases. As before, DVH and PQM metrics were utilized to dosimetrically compare scheduled (iCBCT Verify) and adapted (iCBCT RapidPlan) plans. Timing data was collected to further evaluate the feasibility of integrating this approach within an online adaptive radiotherapy workflow.ResultsModel testing results confirmed the models were capable of producing VMAT plans within a single optimization that were overall improved upon or dosimetrically comparable to original clinical plans. Direct application of RapidPlan on iCBCT datasets produced satisfactory intact prostate and prostate bed plans with generally improved target volume coverage/conformality and rectal sparing relative to iCBCT Verify plans as indicated by DVH values, though bladder metrics were marginally increased on average. Average PQM values for iCBCT RapidPlans were significantly improved compared to iCBCT Verify plans. The average time required [in mm:ss] to generate adapted plans was 06:09 ± 02:06 (intact) and 07:12 ± 01:04 (bed).ConclusionThis study demonstrated the feasibility of leveraging RapidPlan to expeditiously generate adapted VMAT intact prostate and prostate bed plans on iCBCT datasets. In general, adapted plans were dosimetrically improved relative to scheduled plans, emphasizing the practicality of the proposed approach.

AbstractBackgroundAccurate and efficient dose calculation is essential for on‐line adaptive planning in proton therapy. Deep learning (DL) has shown promising dose prediction results in photon therapy. However, there is a scarcity of DL‐based dose prediction methods specifically designed for proton therapy. Successful dose prediction method for proton therapy should account for more challenging dose prediction problems in pencil beam scanning proton therapy (PBSPT) due to its sensitivity to heterogeneities.PurposeTo develop a DL‐based PBSPT dose prediction workflow with high accuracy and balanced complexity to support on‐line adaptive proton therapy clinical decision and subsequent replanning.MethodsPBSPT plans of 103 prostate cancer patients (93 for training and the other 10 for independent testing) and 83 lung cancer patients (73 for training and the other 10 for independent testing) previously treated at our institution were included in the study, each with computed tomography scans (CTs), structure sets, and plan doses calculated by the in‐house developed Monte‐Carlo dose engine (considered as the ground truth in the model training and testing). For the ablation study, we designed three experiments corresponding to the following three methods: (1) Experiment 1, the conventional region of interest (ROI) (composed of targets and organs‐at‐risk [OARs]) method. (2) Experiment 2, the beam mask (generated by raytracing of proton beams) method to improve proton dose prediction. (3) Experiment 3, the sliding window method for the model to focus on local details to further improve proton dose prediction. A fully connected 3D‐Unet was adopted as the backbone. Dose volume histogram (DVH) indices, 3D Gamma passing rates with a criterion of 3%/3 mm/10%, and dice coefficients for the structures enclosed by the iso‐dose lines between the predicted and the ground truth doses were used as the evaluation metrics. The calculation time for each proton dose prediction was recorded to evaluate the method's efficiency.ResultsCompared to the conventional ROI method, the beam mask method improved the agreement of DVH indices for both targets and OARs and the sliding window method further improved the agreement of the DVH indices (for lung cancer, CTV D98 absolute deviation: 0.74 ± 0.18 vs. 0.57 ± 0.21 vs. 0.54 ± 0.15 Gy[RBE], ROI vs. beam mask vs. sliding window methods, respectively). For the 3D Gamma passing rates in the target, OARs, and BODY (outside target and OARs), the beam mask method improved the passing rates in these regions and the sliding window method further improved them (for prostate cancer, targets: 96.93% ± 0.53% vs. 98.88% ± 0.49% vs. 99.97% ± 0.07%, BODY: 86.88% ± 0.74% vs. 93.21% ± 0.56% vs. 95.17% ± 0.59%). A similar trend was also observed for the dice coefficients. This trend was especially remarkable for relatively low prescription isodose lines (for lung cancer, 10% isodose line dice: 0.871 ± 0.027 vs. 0.911 ± 0.023 vs. 0.927 ± 0.017). The dose predictions for all the testing cases were completed within 0.25 s.ConclusionsAn accurate and efficient deep learning‐augmented proton dose prediction framework has been developed for PBSPT, which can predict accurate dose distributions not only inside but also outside ROI efficiently. The framework can potentially further reduce the initial planning and adaptive replanning workload in PBSPT.

AbstractBackgroundDeep learning auto‐segmentation (DLAS) models have been adopted in the clinic; however, they suffer from performance deterioration owing to the clinical practice variability. Some commercial DLAS software provide an incremental retraining function that enables users to train a custom model using their institutional data to account for clinical practice variability.PurposeThis study was performed to evaluate and implement the commercial DLAS software with the incremental retraining function for definitive treatment of patients with prostate cancer in a multi‐user environment.MethodsCT‐based target organs and organs‐at‐risk (OAR) delineation of 215 prostate cancer patients were utilized. The performance of three commercial DLAS software built‐in models was validated with 20 patients. A retrained custom model was developed using 100 patients and evaluated on the remaining data (n = 115). Dice similarity coefficient (DSC), Hausdorff distance (HD), mean surface distance (MSD), and surface DSC (SDSC) were utilized for quantitative evaluation. A multi‐rater qualitative evaluation was blindly performed with a five‐level scale. Visual inspection was performed in consensus and non‐consensus unacceptable cases to identify the failure modes.ResultsThree commercial DLAS vendor built‐in models achieved sub‐optimal performance in 20 patients. The retrained custom model had a mean DSC of 0.82 for prostate, 0.48 for seminal vesicles (SV), and 0.92 for rectum, respectively. This represents a significant improvement over the built‐in model with DSC of 0.73, 0.37, and 0.81 for the corresponding structures. Compared to the acceptance rate of 96.5% and consensus unacceptable rate (i.e., both reviewers rated as unacceptable) of 3.5% achieved by manual contours, the custom model achieved a 91.3% acceptance rate and 8.7% consensus unacceptable rate. The failure modes of retrained custom model were attributed to the following: cystogram (n = 2), hip prosthesis (n = 2), low dose rate brachytherapy seeds (n = 2), air in endorectal balloon(n = 1), non‐iodinated spacer (n = 2), and giant bladder(n = 1).ConclusionThe commercial DLAS software with the incremental retraining function was validated and clinically adopted for prostate patients in a multi‐user environment. AI‐based auto‐delineation of the prostate and OARs is shown to achieve improved physician acceptance, overall clinical utility, and accuracy.

AbstractBackgroundData for proton therapy in high‐risk prostate cancer (HRPC) are limited. Using the Proton Collaborative Group prospective registry, we evaluated outcomes for HRPC patients treated with proton therapy.MethodsA totsl of 605 men with localized HRPC treated with proton therapy from 8/2009 to 3/2019 at nine institutions were selected. Outcomes examined included freedom from progression (FFP), metastasis free survival (MFS), overall survival (OS), and toxicity. Multivariable cox/binomial regression models were used to assess predictors of FFP and toxicity.ResultsMedian age was 71 years. Gleason grade groups 4 (49.4%) and 5 (31.7%) were most common, as were clinical stage T1c (46.1%) and cT2 (41.3%). The median pretreatment prostate specific antigen (PSA) was 9.18 and median International Prostate Symptom Score (IPSS) was 6. Androgen deprivation therapy was given in 63.6%. Median dose was 79.2 GyE in 44 fractions. Pelvic lymph nodes were treated in 58.2% of cases. Pencil beam scanning was used in 54.5%, uniform scanning in 38.8%, and a rectal spacer in 14.2%. At a median followup of 22 months, the 3‐ and 5‐year FFP were 90.7% and 81.4%, respectively. Five‐year MFS and OS were 92.8% and 95.9%, respectively. Independent correlates of FFP included Gleason ≥8, PSA > 10, and cT2 (all p < 0.05). No grade 4 or 5 adverse events were reported. There were 23 (5%) grade 2 and 0 grade 3 gastrointestinal events. Prevalence of late grade 3, late grade 2, acute grade 3, and acute grade 2 genitourinary toxicity was 1.7%, 5.8%, 0%, and 21.8%, respectively. Prevalence of grade 2 and 3 erectile dysfunction at 2 years was 48.4% and 8.4%, respectively.ConclusionsIn the largest series published to date, our results suggest early outcomes using proton therapy for HRPC are encouraging for both safety and efficacy. Further evaluation is needed to determine if an advantage exists to use protons over other radiation techniques in this population.