Yang-Wei Hsieh

Background: Radiotherapy-related complications can impair long-term outcomes in nasopharyngeal carcinoma (NPC), while high-volume literature screening remains a bottleneck in evidence synthesis. Methods: Four databases were searched through January 2026. After deduplication, 4916 records underwent two-stage manual screening, yielding 38 studies. Three Bidirectional Encoder Representations from Transformers (BERT) models, BERT-base, BioBERT, and PubMedBERT, were fine-tuned using stratified five-fold cross-validation, and screening performance was assessed using the area under the receiver operating characteristic curve (ROC-AUC), the area under the precision-recall curve (PR-AUC), and work saved over sampling (WSS) at the 100% and 85% recall thresholds (WSS@100% and WSS@85%).The clinical evidence was evaluated using the Prediction Model Risk of Bias Assessment Tool (PROBAST) and synthesized narratively according to endpoint, discrimination metric, time horizon, unit of analysis, validation level, and uncertainty reporting. Results: At complete recall, BioBERT and PubMedBERT achieved mean WSS values of 97.8% and 97.5%, respectively, compared with 89.7% for BERT-base. Separately, the clinical evidence review identified 39 discrimination estimates: 33 conventional binary ROC-AUCs, three Harrell’s C-indices, and three time-dependent AUCs. Only two studies contributed external-validation estimates and one contributed a temporal-validation estimate. Sixteen estimates lacked a usable measure of uncertainty, and PROBAST rated all studies as having a high overall risk of bias, driven by the analysis domain. Conclusions: Within the retrieved and manually annotated corpus, biomedical pretrained language models showed the potential to reduce the title-and-abstract screening burden while maintaining complete recall. Independently, the clinical evidence map identified promising discrimination but insufficiently robust and transportable evidence for routine clinical implementation.

Objective: This study aims to develop a preliminary explainable image-based auxiliary assessment framework for radiation pneumonitis (RP) by integrating a generative adversarial network (GAN), Gradient-weighted Class Activation Mapping (Grad-CAM), and ensemble learning to improve image-based classification performance, interpretability, and computational efficiency. Methods: Chest CT images from 46 lung cancer patients who underwent VMAT were retrospectively collected, yielding 542 RP and 1857 non-RP images. Images were preprocessed using Otsu-based segmentation and standardized before being input into the RP-GAN for feature extraction. Grad-CAM was applied to qualitatively visualize attention patterns across convolutional layers and guide feature-layer selection, while PCA reduced the extracted features from 12,288 to 730 dimensions. An ensemble stacking classifier combining RF, SVM, KNN, and XGBoost with logistic regression as a meta-learner was constructed. Model performance was evaluated using AUC, accuracy, PPV, NPV, specificity, recall, and F1-score. Results: Grad-CAM highlighted conv2d_4 and conv2d_5 as the most informative layers. PCA reduced training time from 49 min to 49 s with minimal performance loss. In the internal hold-out test set, the ensemble model achieved the highest point estimates for AUC and accuracy among the evaluated classifiers (AUC = 0.921; accuracy = 87.1%). The model identified RP-positive CT slices and provided qualitative visual cues for suspected RP-related regions. Conclusions: The proposed GAN–Grad-CAM ensemble framework showed promising internal classification performance for post-treatment CT-based RP image assessment with substantially improved computational efficiency. Its qualitative visual outputs may support clinical image review, although further external validation and quantitative localization assessment are required before clinical application as an auxiliary image-review tool.

Background/Objectives: This study developed and evaluated a BERT-assisted literature screening workflow to support meta-analyses of postradiotherapy complications in nasopharyngeal carcinoma patients. The aim was to automate key screening steps to improve downstream screening efficiency and consistency, while minimizing time and bias during manual reviews. Materials and Methods: A bidirectional encoder representations from transformers (BERT) model was integrated into a standard systematic review pipeline for studies on postradiotherapy complications in nasopharyngeal carcinoma. The workflow combined automated BERT-based classification with manual verification and followed PRISMA and PICOS guidelines for literature identification, screening, and eligibility assessment. Model training involved hyperparameter tuning and comparison of different optimizers to maximize screening performance against a manually curated reference set, with particular attention to discrimination (AUC) and processing time. Results: From an initial corpus of 6496 records, the combined automated and manual workflow identified 23 eligible studies for meta-analysis. The included studies showed substantial heterogeneity (I² = 86.85%), supporting the use of a random-effects model to pool outcomes. The BERT model optimized with an Adagrad optimizer achieved an AUC of 0.77 for relevant-study classification and reduced screening time to 1142 s. To demonstrate the workflow’s utility, a downstream meta-analysis was conducted using the identified studies. As a downstream application based on the identified studies, a quantitative synthesis was conducted, in which (meta-analysis of the 23 included studies), a random forest model—evaluated across those studies—achieved an AUC of 0.92 under a fixed-effect analysis for predicting postradiotherapy complications. Conclusions: Integrating BERT into the literature screening phase of meta-analysis for postradiotherapy nasopharyngeal carcinoma complications markedly improved screening efficiency while maintaining acceptable classification performance. This workflow demonstrates the feasibility of transformer-based assistance for systematic reviews and provides a foundation for developing disease-specific, AI-augmented evidence synthesis pipelines in oncology.

Background: Conventional cone‐beam computed tomography (CBCT) systems are limited by suboptimal image quality, inaccurate Hounsfield unit (HU) calibration, and reduced reliability for dose calculation. HyperSight CBCT on the Halcyon platform offers improved HU accuracy, expanded field‐of‐view (FOV), and enhanced image quality. Purpose: This study aimed to assess the dosimetric accuracy of treatment planning using HyperSight CBCT through phantom‐based dose verification. Methods: This study included three institutions equipped with the HyperSight imaging system on the Halcyon platform, with all procedures performed after acceptance testing and calibration. Each institution generated HU‐to‐density calibration curves using computed tomography (CT) scanners and standardized phantoms, and corresponding CBCT for planning (CBCTp) scans were also acquired. Additional CBCTp scans were acquired using a consistent phantom model (062 M) across the three institutions. Reference treatment plans were created on CT images and transferred to CBCTp and CBCT datasets for dose recalculation using identical parameters. Dosimetric assessment included gamma analysis and comparisons of DVH‐based dosimetric metrics for relevant regions of interest (ROIs). End‐to‐end testing with an anthropomorphic phantom was performed using ion chamber measurements and film dosimetry at brain, bone, and thorax locations. Results: HU‐to‐density curves showed consistent behavior across institutions, with larger variability only at higher densities. CBCTp calibrations agreed well with vendor references. DVH‐based dosimetric metrics showed differences generally within 1% for both CBCTp and CBCT when compared with CT. Across institutions, gamma analysis of both CBCTp and CBCT yielded high passing rates (≥ 98.5% at 3%/2 mm). End‐to‐end testing with film dosimetry showed that CBCT‐based plans agreed with measured doses within ± 4%, while CT‐based plans were within ± 3%. Ion chamber measurements showed all dose differences within ± 2.3%, with both CBCTp and CBCT within ± 1.0% of CT. Conclusions: HyperSight CBCT provides accurate dose calculations when properly calibrated. Phantom‐based validation demonstrated sub‐2% deviations and strong agreement with CT, supporting its clinical use in adaptive radiotherapy.

Purpose: This study develops a predictive model for radiation pneumonitis (RP) risk in lung cancer patients after volume-modulated arc therapy (VMAT) that leverages high-dimensional dosiomics and dose–volume histogram (DVH) features within IDLSS (incremental-dose interval-based lung subregion) lung subregions. Methods: We retrospectively analyzed data from 136 lung cancer patients treated with VMAT between 2015 and 2022, including 39 patients who developed RP greater than Grade 2. Using the IDLSS method, seven regions of interest (ROIs), including the Planning Target Volume (PTV), normal lung, and five subdivided lung areas, were delineated on pretreatment Computed Tomography (CT) images. DVH, radiomics, and dosiomics features were extracted from these ROIs and organized into nine distinct feature sets. A comprehensive pipeline was applied, integrating IDLSS-defined lung subregions, high-dimensional dosiomics features, LASSO-based feature selection, and SMOTE oversampling to address class imbalance in the training data. Logistic regression, random forest, and feedforward neural networks were constructed and optimized via tenfold cross-validation. Model performance across different feature sets was evaluated via the average AUC, F1 score, and other performance metrics. Results: LASSO regression revealed that BMI and volume within the 5–10 Gy and 10–20 Gy lung subregions were significant predictors of RP. The performance evaluation demonstrated that the dosiomics features consistently outperformed the DVH features across the models. Combining radiomics and dosiomics achieved the highest predictive accuracy (AUC = 0.91, ACC = 0.89, NPV = 0.95, PPV = 0.78, F1 score = 0.82, sensitivity = 0.88, specificity = 0.90). Applying SMOTE during training significantly improved sensitivity without compromising specificity, confirming the value of balancing strategies in enhancing model performance. Incorporating all the features together did not provide additional performance gains. Conclusions: Integrating radiomics and dosiomics features extracted from IDLSS-defined lung subregions significantly enhances the ability to predict RP after VMAT, surpassing traditional DVH metrics. The substantial contribution of dosiomics features highlights the importance of spatial dose heterogeneity in RP risk assessment.

Introduction: This study aims to develop an individualized risk prediction model for radiation pneumonitis (RP) based on unsupervised image feature learning. A deep convolutional generative adversarial network (DCGAN) was utilized to automatically extract features from computed tomography (CT) images. Methods: A retrospective analysis was conducted on 180 lung cancer patients treated with volumetric modulated arc therapy (VMAT) at Kaohsiung Veterans General Hospital between 2015 and 2022. To mitigate clinical sample size limitations, rotation-based augmentation was employed to expand the training dataset. The pretreatment CT images were processed into three input configurations: whole-lung, V _5Gy dose regions, and V _20Gy dose regions. An unsupervised feature extraction model, designated RP-GAN, was constructed to capture latent representations associated with RP risk. High-dimensional features were refined via least absolute shrinkage and selection operator (LASSO) and integrated into a stacking ensemble learning framework (including RF, SVM, KNN, XGBoost, and LR). Model stability and generalization were validated through 10-fold cross-validation alongside an independent test set, while clinical interpretability was ensured using Grad-CAM and LIME. Results: The whole-lung input model demonstrated superior performance, achieving an AUC of 0.856 and an accuracy of 0.861, with a recall of 0.778. In contrast, models restricted to V _20Gy dose regions showed a significant decline in sensitivity, with the recall decreasing to 0.273. XAI visualization confirmed that the model focused not only on the tumor bed but also on the peritumoral parenchyma and contralateral lung. Conclusion: The proposed RP-GAN architecture effectively captures subtle textural changes across the whole lung microenvironment without requiring manual annotations. This framework provides a robust tool for individualized RP risk assessment, facilitating the optimization of radiation therapy plans.

Purpose: To improve the prediction accuracy of radiation dermatitis (RD) in breast cancer patients undergoing volumetric modulated arc therapy (VMAT), we developed a hybrid artificial intelligence (AI) model that integrates deep learning radiomics (DLR), handcrafted radiomics (HCR), clinical features, and dose–volume histogram (DVH) parameters, aiming to enhance the early identification of high-risk individuals and support personalized prevention strategies. Methods: A retrospective cohort of 156 breast cancer patients treated with VMAT at Kaohsiung Veterans General Hospital (2018–2023) was analyzed; 148 patients were eligible after exclusions, with RD graded according to the RTOG criteria. Clinical variables and 12 DVH indices were collected, while HCR features were extracted via PyRadiomics. DLR features were derived from a pretrained VGG16 network across four input designs: original CT images (DLR_Original), a 5 mm subcutaneous region (DLR_Skin5mm), a planning target volume with a 100% prescription dose (DLR_PTV100%), and a subcutaneous region receiving ≥ 5 Gy (DLR_V5Gy). The features were preselected via ANOVA (p < 0.05), followed by Boruta–SHAP refinement across 11 feature sets. Predictive models were built via logistic regression, random forest, gradient boosting decision tree, and stacking ensemble (SE) methods. Explainability was assessed via SHapley Additive exPlanations (SHAPs) and gradient-weighted class activation mapping (Grad-CAM). Results: Among the 148 patients, 49 (33%) developed Grade ≥ 2 RD. The DLR models outperformed the HCR models (AUC = 0.72 vs. 0.66). The best performance was achieved with DLR_V5Gy + clinical + DVH features, yielding an AUC = 0.76, recall = 0.68, and F1 score = 0.60. SE consistently surpassed single classifiers. SHAP identified convolutional DLR features as the strongest predictors, whereas Grad-CAM focused attention on subcutaneous high-dose regions, which was consistent with the clinical RD distribution. Conclusions: The proposed hybrid AI framework, which integrates DLR, clinical, and DVH features, provides accurate and explainable predictions of Grade ≥ 2 RD after VMAT in breast cancer patients. By combining ensemble learning with XAI methods, the model offers reliable high-risk stratification and potential clinical utility for personalized treatment planning.

Purpose: This study aimed to develop a precise predictive model to assess the risk of radiation pneumonitis (RP) and three-year survival in patients with non-small cell lung cancer (NSCLC) following volumetric modulated arc therapy (VMAT). Radiomics features, ensemble stacking, and explainable artificial intelligence (XAI) were integrated to enhance predictive performance and clinical interpretability. Materials and Methods: A retrospective cohort of 221 NSCLC patients treated with VMAT at Kaohsiung Veterans General Hospital between 2013 and 2023 was analyzed, including 168 patients for RP prediction (47 with ≥grade 2 RP) and 118 patients for survival prediction (34 deaths). Clinical variables, dose–volume histogram (DVH) parameters, and radiomic features (original, Laplacian of Gaussian [LoG], and wavelet filtered) were extracted. ANOVA was used for initial feature reduction, followed by LASSO and Boruta-SHAP for feature selection, which formed 10 feature subsets. The data were divided at an 8:2 ratio into training and testing sets, with SMOTE balancing and 10-fold cross-validation for parameter optimization. Six models—logistic regression (LR), random forest (RF), support vector machine (SVM), k-nearest neighbors (KNN), XGBoost, and Ensemble Stacking—were evaluated in terms of the AUC, accuracy (ACC), negative predictive value (NPV), precision, and F1 score. SHAP analysis was applied to interpret feature contributions. Results: For RP prediction, the LASSO-selected radiomic subset (FR) combined with Ensemble Stacking achieved optimal performance (AUC 0.91, ACC 0.89), with SHAP identifying V40 Firstorder_Min as the most influential feature. For survival prediction, the FR subset yielded an AUC of 0.97, an ACC of 0.92, and an NPV of 1.00, with V10 Wavelet Firstorder_Min as the top contributor. The multimodal subset (FC+R) also performed strongly, achieving an AUC of 0.91 for RP and 0.96 for survival. Conclusions: This study demonstrated the superior performance of radiomics combined with Ensemble Stacking and XAI for the prediction of RP and survival following VMAT in patients with NSCLC. SHAP-based interpretation enhances transparency and clinical trust, offering a robust foundation for personalized radiotherapy and precision medicine.

Background: Growing patients with nasopharyngeal carcinoma (NPC) were treated with intensity-modulated proton therapy (IMPT). However, a high probability of severe acute radiation dermatitis (ARD) was observed. The objective of the study is to investigate the dosimetric parameters related to ARD for NPC patients treated with IMPT. Methods: Sixty-two patients with newly diagnosed NPC were analyzed. The ARD was recorded based on the criteria of Common Terminology Criteria for Adverse Events version 4.0. Logistic regression model was performed to identify the clinical and dosimetric parameters related to ARD. Receiver operating characteristic (ROC) curve analysis and the area under the curve (AUC) were used to evaluate the performance of the models. Results: The maximum ARD grade was 1, 2, and 3 in 27 (43.5%), 26 (42.0%), and 9 (14.5%) of the patients, respectively. Statistically significant differences (p < 0.01) in average volume to skin 5 mm with the respective doses were observed in the range 54–62 Cobalt Gray Equivalent (CGE) for grade 2 and 3 versus grade 1 ARD. Smoking habit and N2-N3 status were identified as significant predictors to develop grade 2 and 3 ARD in clinical model, and V58CGE to skin 5 mm as an independent predictor in dosimetric model. After adding the variable of V58CGE to the metric incorporating two parameters of smoking habit and N status, the AUC value of the metric increases from 0.78 (0.66–0.90) to 0.82 (0.72–0.93). The most appropriate cut-off value of V58CGE to skin 5 mm as determined by ROC curve was 5.0 cm³, with a predicted probability of 54% to develop grade 2 and 3 ARD. Conclusion: The dosimetric parameter of V58CGE to skin 5 mm < 5.0 cm³ could be used as a constraint in treatment planning for NPC patients treated by IMPT.