Dongrong Yang

All Publications

• Reference-Free large language model agents for physician-guided radiotherapy treatment planning. Medical physics Yang, D., Wu, X., Xie, Y., Li, X., Wu, Q., Wu, Q. J., Sheng, Y. 2026; 53 (8): e70592

  Abstract

  Large language models (LLMs) have recently demonstrated exceptional capabilities, offering the potential to streamline workflows and enhance efficiency across diverse tasks. However, their application in domain-specific areas, such as radiotherapy treatment planning, remains challenging due to the lack of publicly available, specialized knowledge required for these tasks.To investigate the critical architectural and functional components necessary to adapt an off-the-shelf large language model into a clinically viable planning agent for inverse treatment planning in intensity-modulated radiation therapy (IMRT).Twenty head-and-neck (HN) cancer patients who received IMRT at our institution were retrospectively collected under IRB approval. The LLM agent was implemented to directly interact with the clinical treatment planning system (TPS) to iteratively extract intermediate plan states and propose new constraint values to guide inverse optimization. Its decision-making was informed by real-time plan evaluations and prior optimization outcomes, enabling adaptive refinement of planning strategies across iterations. The agent was equipped with three core capabilities: comprehension of clinical planning objectives, contextual understanding of the optimization environment, and arithmetic proficiency for quantitative reasoning. Treatment planning was conducted in a reference-free inference setting, wherein the LLM operated without prior exposure to manually generated treatment plans and without any fine-tuning or task-specific training. LLM-generated treatment plans were compared with clinically approved plans created by certified dosimetrists. Key dosimetric endpoints were evaluated and statistically analyzed. Paired comparisons between LLM-generated and clinical plans were performed using the Wilcoxon signed-rank test.LLM-generated plans achieved comparable organ-at-risk (OAR) sparing relative to clinical plans, while demonstrating improved hot spot control (Dmax: 106.5% vs. 108.8%, p < 0.05) and superior conformity (conformity index: 1.18 vs. 1.39, p < 0.05 for boost PTV; 1.82 vs. 1.88, p = 0.47 for primary PTV).This study demonstrates the feasibility of a reference-free, LLM-driven workflow for automated IMRT treatment planning in a commercial TPS. The proposed approach provides a generalizable solution that could reduce planning variability and support broader adoption of AI-based planning strategies.

  View details for DOI 10.1002/mp.70592

  View details for PubMedID 42535478

• Illusion of convergence: Search space geometry in radiotherapy treatment plan optimization. Medical physics Wu, X., Yang, D., Sheng, Y., Ge, Y., Wu, Q. J., Wu, Q. 2026; 53 (1): e70209

  Abstract

  Treatment plan optimization is foundational for radiotherapy. However, the geometry of the underlying search space where feasible solutions reside remains poorly characterized. Understanding this search space is crucial as it: (1) Exposes hidden limitations in objective functions; (2) Reveals the nature of convergence; and (3) Informs the choice of optimization algorithms and initialization strategies.To characterize the geometry of the search space, examine the nature of convergence, and provide theoretical guidance for empirical rules in clinical practice.This study examines fluence map optimization (FMO) for intensity-modulated radiotherapy (IMRT) using a L- BFGS based framework. Hessian matrix of objective function was derived, and its eigenvalues were determined by the voxels actively contributing to the objective and the singular value decomposition (SVD) of the dose-deposition matrix. Numerical analysis includes estimating dominant eigenvalues and corresponding eigenvectors. Perturbation analysis was conducted along individual beamlet intensity and principal eigenvectors for visualization. The primary focus was on quadratic dose-volume objectives (DVOs), with extensions to generalized equivalent uniform dose (gEUD).Eigenvalue spectra showed highly anisotropic curvature near the optimum, dominated by broad flat regions with minimal sensitivity in most directions. Perturbation plots revealed that many beamlets were either insensitive or exhibited step-like, discontinuous behavior. Convergence often occurred in such flat regions, producing an "illusion of convergence" that did not guarantee the lowest basin. Optimization track comparisons showed that second-order methods, by incorporating curvature, advanced more efficiently than first-order methods.For DVO-based FMO, search space is dominated by flat plateaus at different altitudes, making convergence sensitive to both optimizer and initialization. The findings explain why clinical plans are often locally but not globally optimal, while still acceptable in quality. The analysis provides a theoretical foundation for longstanding empirical practices and offers guidance on optimizer selection and initialization strategies in treatment plan optimization.

  View details for DOI 10.1002/mp.70209

  View details for PubMedID 41423572

• Enhanced 4D diffusion-weighted PROPELLER echo-planar imaging with collaborative blade reconstruction. Physics in medicine and biology Wang, L., Li, T., Liu, C., Yang, D., Cai, J., Chang, H. C. 2025; 70 (23)

  Abstract

  Objective.Four-dimensional diffusion-weighted imaging (4D-DWI) offers enhanced soft-tissue contrast for image-guided radiotherapy (IGRT) compared to conventional 4D-magnetic resonance imaging. However, its clinical application has been hindered by geometric distortion or prolonged acquisition times. This study introduces an improved respiratory-correlated 4D-DWI technique to address these critical limitations.Approach.The proposed 4D-DW-PROP-EPI-JBCR technique consists of four stages: data acquisition, coil sensitivity and phase error maps estimation, retrospective data sorting, and a joint-blade collaborative reconstruction (JBCR) step. We evaluated its performance against the previous 4D-DW-PROPELLER-EPI using both simulations andin vivoexperiments. Evaluation metrics includedk-space uniformity requirements, point spread function (PSF) quality, minimum required sampling factor (SF), image quality, geometric fidelity, and acquisition time. Image quality was assessed using peak signal-to-noise ratio (PSNR), structural similarity (SSIM), high-frequency error norm (HFEN), and signal-to-noise ratio (SNR); geometric fidelity was evaluated via Dice similarity coefficient (DSC) and Hausdorff distance (HD). Pairedt-tests were used for metric comparisons, with statistical significance set atP< 0.05.Main results.Compared to 4D-DW-PROPELLER-EPI, the proposed 4D-DW-PROP-EPI-JBCR eliminated the requirement for strictk-space uniformity while achieving enhanced PSFs with fewer acquisitions. Additionally, 4D-DW-PROP-EPI-JBCR yielded significant improvements in image quality compared to 4D-DW-PROPELLER-EPI across all key metrics: SSIM (0.83-0.92 vs 0.81-0.90,P< 0.05), PSNR (20.56-25.20 vs 20.24-24.48,P< 0.05), HFEN (0.49-0.678 vs 0.52-0.80, P < 0.05), and SNR (31.1-32.6 dB vs 28.6-29.5 dB,P< 0.05). It also exhibited excellent geometric fidelity, as evidenced by DSC (0.90 ± 0.03) and HD (2.37 ± 1.12). Notably, these performance gains were achieved with an SF of 1.9 per diffusion direction, compared to the SF of 2.4 for 4D-DW-PROPELLER-EPI.Significance.By enabling distortion-free respiratory-correlated 4D-DWI and reducing SF by more than 20% compared to 4D-DW-PROPELLER-EPI, the 4D-DW-PROP-EPI-JBCR technique markedly improves the clinical utility of 4D-DWI for abdominal IGRT.

  View details for DOI 10.1088/1361-6560/ae1fce

  View details for PubMedID 41237423

• Foresight planning: Radiotherapy plan optimization via self-supervised model predictive control. Medical physics Yang, D., Wu, X., Xie, Y., Li, X., Wu, Q., Wu, Q. J., Sheng, Y. 2025; 52 (11): e70132

  Abstract

  Treatment planning for intensity-modulated radiation therapy (IMRT) and volumetric modulated arc therapy (VMAT) relies on meticulous operation of the inverse optimization, a complex and iterative process of adjusting dose-volume objectives to achieve an optimal dose distribution. This is a trial-and-error process due to the black box nature of the optimization engine.To propose a novel foresight planning strategy that models the inverse optimization process, providing direct guidance to streamline optimization toward the desired dose distribution with efficiency and consistency.The proposed strategy involves two stages. First, a Deep-Dose-Predictive (DDP) model was trained to predict the dose response based on historical plan states and dose-volume objective (DVO) adjustments. The training of the DDP model mirrored that of a clinical planner, during which it gained intelligence in understanding dose-response principles by witnessing extensive plan state transition history. The training dataset was generated using Monte Carlo sampling without any intervention. In the second stage, the trained DDP model was employed for automatic DVO adjustments via model predictive control. By forecasting future plan states based on potential DVO modifications, the model assessed plan quality using a score function, which evaluated predicted dose responses across all possible adjustments. The adjustment maximizing the score function was selected, enabling the strategy to adapt to specific clinical priorities through score function weighting, without requiring model retraining. The feasibility of the method was validated in head-and-neck cancer IMRT. A total of 40 cases were used to collect plan state transitions for model training, while an additional 40 patients were utilized for model evaluation. The proposed framework was tasked to generate plans with different parotid-sparing priorities (i.e., bilateral and unilateral sparing) based on patient-specific conditions.The automatically generated plans demonstrate clinically comparable quality for both bilateral and unilateral sparing cases. For bilateral sparing cases, the automated plans achieve non-inferior organ-at-risk (OAR) sparing while exhibiting superior conformity indices-1.62 and 1.15 for the primary PTV and boost PTV, respectively, compared to 1.79 and 1.47 for clinical plans. Similarly, for unilateral cases, the automated plans maintain non-inferior OAR sparing with improved conformity indices of 1.78 and 1.14, compared to 2.07 and 1.42 for clinical plans.The proposed strategy effectively automates radiation therapy planning, achieving clinically comparable plan quality with improved efficiency and adaptability. This approach introduces a transformative perspective to automating treatment planning, paving the way for more intelligent and flexible AI solutions for radiation therapy treatment planning.

  View details for DOI 10.1002/mp.70132

  View details for PubMedID 41188012

• A technique for achieving arbitrary machine dose rates via MLC leaf motion modulation using fixed-gantry IMRT delivery mode. Medical physics Wu, X., Yang, D., Wu, Q. J., Sheng, Y., Rodrigues, A. E., Kowalski, J. P., Wu, Q. 2025; 52 (6): 3595-3610

  Abstract

  The medical linear accelerator (Linac) typically operates at a constant but discrete dose rate for both 3D conformal fields and intensity modulation radiotherapy (IMRT) fields. However, in certain clinical and radiobiological scenarios, such as total body irradiation (TBI), an arbitrary dose rate outside the preset values may be desired.This study aims to achieve arbitrary machine dose rates in photon delivery through modulation of MLC leaf motion, providing flexibility for specific needs.The desired effective dose rate is achieved by precisely programming the motion of unused MLC leaves outside the field boundary set by the jaws, while keeping the MLC leaves within the jaw boundary same as originally planned. Between control points, the leaf speed is adjusted by multiplying the permissible limit by the ratio of the current dose rate to the desired value. During delivery, the dose rate is reduced so that the MLC leaves arrive at the expected position for the expected monitor units (MU). Two schemes were designed for different application scenarios and successfully delivered and verified. A circular 3D conformal field, a 40 cm × 40 cm TBI field, and a head-and-neck (H&N) IMRT DMLC field were modified for dose rates between 600 and 1 MU/min. These plans were successfully delivered on Varian TrueBeam, and Clinac Linacs with modulated dose rates. Trajectory log analysis and portal dosimetry QA (PDQA) were utilized to verify the accuracy.Both schemes achieved stable and accurate dose rates, with mean deviations from the desired values remaining below 0.5% across the range from 600 to 1 MU/min. Fluctuations, represented by relative standard deviations, increased monotonically as the desired dose rate decreased. For clinically relevant dose rates above 100 MU/min, fluctuations remained below 0.5%. At lower dose rates, the thresholds for fluctuations reaching 1% and 10% were approximately 50 and 7 MU/min, respectively. Trajectory log file analysis confirmed agreement between planned and actual values of various treatment parameters, such as leaf position, speed, and MU. PDQA analysis showed 99% gamma pass rate with criteria of 3%/2 mm. For Scheme 1 with step-and-shoot, a 50% increase in delivery time due to the temporary beam-hold during step-and-shoot was both expected and observed.A technique has been developed and validated to achieve arbitrary machine dose rates in photon delivery through the modulation of MLC leaf motion.

  View details for DOI 10.1002/mp.17835

  View details for PubMedID 40280879

• A machine learning toolkit assisted approach for IMRT fluence map optimization: feasibility and advantages. Biomedical physics & engineering express Wu, X., Yang, D., Sheng, Y., Wu, Q. J., Wu, Q. 2025; 11 (3)

  Abstract

  Purpose. Traditional machine learning (ML) and deep learning (DL) applications in treatment planning rely on complex model architectures and large, high-quality training datasets. However, they cannot fully replace the conventional optimization process. This study presents a novel application of ML in treatment planning where established ML/DL toolkits are directly applied to treatment plan optimization.Materials and Methods. A one-layer network was designed based on the dose deposition matrix and implemented in PyTorch's L-BFGS optimizer with GPU acceleration. The classical steepest descent optimizer was selected as a reference for comparison. Both optimizers utilized identical inputs and objective functions to ensure a fair comparison. DVH- and gEUD-based objectives were implemented in standard quadratic forms. Standard uniform and 1,000 random initializations were used to test optimizer's search ability under different starting conditions for prostate and head-and-neck cases.Results. The MLT-assisted framework demonstrated comparable or superior plan quality to classical optimization by achieving lower objective values, improved DVHs and capturing finer modulation details in fluence maps. For gEUD-based optimization, it effectively explored beam weight elevations that classical optimization could only reach with stricter convergence criteria and many more iterations. The quality differences primarily stemmed from convergence speed. The MLT-assisted framework required significantly fewer evaluations and iterations to achieve similar or better results. Optimization on random initial maps further demonstrated that it was more robust and less likely to be trapped. It does not require stricter convergence criteria or extended runs to reach high-quality optima, making it more efficient and reliable.Conclusion. This framework leverages ML toolkits in a novel way, enabling faster convergence, greater robustness and handling of complex constraints. As the first study of its kind, it establishes MLT-assisted optimization as a viable and effective alternative to classical methods.

  View details for DOI 10.1088/2057-1976/adcaca

  View details for PubMedID 40203852

• Reinforcement learning-driven automated head and neck simultaneous integrated boost (SIB) radiation therapy: flexible treatment planning aligned with clinical preferences. Physics in medicine and biology Yang, D., Wu, X., Li, X., Xie, Y., Wu, Q., Wu, Q. J., Sheng, Y. 2025; 70 (8)

  Abstract

  Objective.Head-and-neck simultaneous integrated boost (SIB) treatment planning using intensity modulated radiation therapy is particularly challenging due to the proximity to organs-at-risk. Depending on the specific clinical conditions, different parotid-sparing strategies are utilized to preserve parotid function without compromising local tumor control. Clinically this is typically done with attending's directive or via trial-and-error comparison with different sparing tradeoffs. To streamline this process, we proposed a deep reinforcement learning (DRL)-based framework that automatically generates treatment plans with flexibility to adapt to clinical preferences.Approach.A preference-encoded DRL (PEDRL) framework was developed to self-interact with the clinical treatment planning system and dynamically adjust objective constraints in the inverse optimization space. It was powered by the discrete soft actor-critic algorithm with a multi-layer perceptron architecture. The agent interprets intermediate plan status and iteratively modifies objective constraint values in a human-like fashion. By encoding parotid-sparing preferences within the state space, the agent autonomously adapts the sparing strategy to achieve optimal plan quality based on clinical priorities. The agent was trained through iterative treatment plan generation using 40 cases and subsequently tested on additional 44 patients, with generated plans compared to clinical plans.Main results.The PEDRL-generated plans demonstrated comparable performance across all dosimetric evaluation metrics for both bilateral and unilateral sparing cases in the test set. For bilateral cases, the mean value of the parotid median dose was 18.82 Gy (left) and 19.61 Gy (right), compared to 19.31 Gy (left) and 19.12 Gy (right) in the clinical plans. In unilateral sparing cases, the mean value of the spared parotid median dose was 19.92 Gy in the PEDRL-generated plans, compared to 17.16 Gy in the clinical plans. Significance.The proposed novel automated treatment planning framework efficiently generates SIB treatment plans tailored to clinical preferences, demonstrating both effectiveness and adaptability.

  View details for DOI 10.1088/1361-6560/adcb84

  View details for PubMedID 40209749

• Breast radiation therapy fluence painting with multi-agent deep reinforcement learning. Medical physics Dongrong, Y., Xinyi, L., Sua, Y., Rachel, B., Susan, M., Sarah, S., Paul, S., Q Jackie, W., Yang, S. 2025; 52 (4): 2015-2024

  Abstract

  The electronic compensation (ECOMP) technique for breast radiation therapy provides excellent dose conformity and homogeneity. However, the manual fluence painting process presents a challenge for efficient clinical operation.To facilitate the clinical treatment planning automation of breast radiation therapy, we utilized reinforcement learning (RL) to develop an auto-planning tool that iteratively edits the fluence maps under the guidance of clinically relevant objectives.With institutional review board (IRB) approval, 70 patients treated with 6MV tangential photon beams with ECOMP technique were retrospectively collected and included in this study (20/50 for training/testing). Each pixel in the fluence map was assigned a reinforcement learning agent to perform independent action. Beam-eye-view projected dose profiles were generated to form state information as the input of the RL network. By predicting the Q value, pixel-wise actions were selected to modify specific pixel value in the fluence maps to improve overall plan quality. After dose calculation, reward signal calculated from the variation of target coverage and dose homogeneity was fed back to the RL framework and used to update network parameters. The RL generated plans were evaluated with dose distribution and dosimetric endpoints (i.e., Breast PTV V90%, Breast PTV V95%, Breast PTV V105%, Lung V20 Gy, Heart V5 Gy, Dmax) and compared with clinical plans.The RL agent took around 90 s to generate a ECOMP treatment plan. The RL plans exhibited plan quality comparable to clinical plans in terms of isodose distribution and dosimetric endpoints. The mean Breast PTV V95%, Breast PTV V105% of RL plans are 77.759 % ( ± 8.904 % ) $77.759{\mathrm{\ \% }}( { \pm 8.904{\mathrm{\ \% }}} )$ and 8.522 cc ( ± 11.469 cc ) $8.522{\mathrm{\ cc\ }}( { \pm 11.469{\mathrm{\ cc}}} )$ , compared to 78.568 % ( ± 9.094 % ) $78.568{\mathrm{\ \% }}( { \pm 9.094{\mathrm{\ \% }}} )$ and 34.298 cc ( ± 36.297 cc ) $34.298\ {\mathrm{cc}}\ ( { \pm 36.297{\mathrm{\ cc}}} )$ cc of clinical plans.The developed RL framework efficiently generates breast ECOMP plans with clinical acceptable plan quality.

  View details for DOI 10.1002/mp.17615

  View details for PubMedID 39853548

• Clinical commissioning and introduction of an in-house artificial intelligence (AI) platform for automated head and neck intensity modulated radiation therapy (IMRT) treatment planning. Journal of applied clinical medical physics Li, X., Sheng, Y., Wu, Q. J., Ge, Y., Brizel, D. M., Mowery, Y. M., Yang, D., Yin, F. F., Wu, Q. 2025; 26 (1): e14558

  Abstract

  To describe the clinical commissioning of an in-house artificial intelligence (AI) treatment planning platform for head-and-neck (HN) Intensity Modulated Radiation Therapy (IMRT).The AI planning platform has three components: (1) a graphical user interface (GUI) is built within the framework of a commercial treatment planning system (TPS). The GUI allows AI models to run remotely on a designated workstation configured with GPU acceleration. (2) A template plan is automatically prepared involving both clinical and AI considerations, which include contour evaluation, isocenter placement, and beam/collimator jaw placement. (3) A well-orchestrated suite of AI models predicts optimal fluence maps, which are imported into TPS for dose calculation followed by an optional automatic fine-tuning. Six AI models provide flexible tradeoffs in parotid sparing and Planning Target Volume (PTV)-organ-at-risk (OAR) preferences. Planners could examine the plan dose distribution and make further modifications as clinically needed. The performance of the AI plans was compared to the corresponding clinical plans.The average plan generation time including manual operations was 10-15  min per case, with each AI model prediction taking ∼1 s. The six AI plans form a wide range of tradeoff choices between left and right parotids and between PTV and OARs compared with corresponding clinical plans, which correctly reflected their tradeoff designs.The in-house AI IMRT treatment planning platform was developed and is available for clinical use at our institution. The process demonstrates outstanding performance and robustness of the AI platform and provides sufficient validation.

  View details for DOI 10.1002/acm2.14558

  View details for PubMedID 39503512

  View details for PubMedCentralID PMC11712748

• Automated treatment planning with deep reinforcement learning for head-and-neck (HN) cancer intensity modulated radiation therapy (IMRT). Physics in medicine and biology Yang, D., Wu, X., Li, X., Mansfield, R., Xie, Y., Wu, Q., Jackie Wu, Q., Sheng, Y. 2024; 70 (1)

  Abstract

  Purpose.To develop a deep reinforcement learning (DRL) agent to self-interact with the treatment planning system to automatically generate intensity modulated radiation therapy (IMRT) treatment plans for head-and-neck (HN) cancer with consistent organ-at-risk (OAR) sparing performance.Methods.With IRB approval, one hundred and twenty HN patients receiving IMRT were included. The DRL agent was trained with 20 patients. During each inverse optimization process, the intermediate dosimetric endpoints' values, dose volume constraints' values and structure objective function losses were collected as the DRLstates. By adjusting the objective constraints asactions, the agent learned to seek optimal rewards by balancing OAR sparing and planning target volume (PTV) coverage. Reward computed from current dosimetric endpoints and clinical objectives were sent back to the agent to update action policy during model training. The trained agent was evaluated with the rest 100 patients.Results.The DRL agent was able to generate a clinically acceptable IMRT plan within12.4±3.1min without human intervention. DRL plans showed lower PTV maximum dose (109.2%) compared to clinical plans (112.4%) (p< .05). Average median dose of left parotid, right parotid, oral cavity, larynx, pharynx of DRL plans were 15.6 Gy, 12.2 Gy, 25.7 Gy, 27.3 Gy and 32.1 Gy respectively, comparable to 17.1 Gy, 15.7 Gy, 24.4 Gy, 23.7 Gy and 35.5 Gy of corresponding clinical plans. The maximum dose of cord + 5 mm, brainstem and mandible were also comparable between the two groups. In addition, DRL plans demonstrated reduced variability, as evidenced by smaller 95% confidence intervals. The total MU of the DRL plans was 1611 vs 1870 (p< .05) of clinical plans. The results signaled the DRL's consistent planning strategy compared to the planners' occasional back-and-forth decision-making during planning.Conclusion.The proposed DRL agent is capable of efficiently generating HN IMRT plans with consistent quality.

  View details for DOI 10.1088/1361-6560/ad965d

  View details for PubMedID 39577088

• Understanding and modeling human-AI interaction of artificial intelligence tool in radiation oncology clinic using deep neural network: a feasibility study using three year prospective data. Physics in medicine and biology Yang, D., Murr, C., Li, X., Yoo, S., Blitzblau, R., McDuff, S., Stephens, S., Wu, Q. J., Wu, Q., Sheng, Y. 2024; 69 (22)

  Abstract

  Objective.Artificial intelligence (AI) based treatment planning tools are being implemented in clinic. However, human interactions with such AI tools are rarely analyzed. This study aims to comprehend human planner's interaction with the AI planning tool and incorporate the analysis to improve the existing AI tool.Approach.An in-house AI tool for whole breast radiation therapy planning was deployed in our institution since 2019, among which 522 patients were included in this study. The AI tool automatically generates fluence maps of the tangential beams to create anAI plan. Human planner makes fluence edits deemed necessary and after attending physician approval for treatment, it is recorded asfinal plan. Manual modification value maps were collected, which is the difference between theAI-planand thefinal plan. Subsequently, a human-AI interaction (HAI) model using full scale connected U-Net was trained to learn such interactions and perform plan enhancements. The trained HAI model automatically modifies theAI planto generate AI-modified plans (AI-m plan), simulating human editing. Its performance was evaluated against originalAI-planandfinal plan. Main results. AI-m planshowed statistically significant improvement in hotspot control over theAI plan, with an average of 25.2cc volume reduction in breast V105% (p= 0.011) and 0.805% decrease in Dmax (p< .001). It also maintained the same planning target volume (PTV) coverage as thefinal plan, demonstrating the model has captured the clinic focus of improving PTV hot spots without degrading coverage.Significance.The proposed HAI model has demonstrated capability of further enhancing theAI planvia modeling human-AI tool interactions. This study shows analysis of human interaction with the AI planning tool is a significant step to improve the AI tool.

  View details for DOI 10.1088/1361-6560/ad8e29

  View details for PubMedID 39488080

• Dynamic Chest Radiograph Simulation Technique with Deep Convolutional Neural Networks: A Proof-of-Concept Study. Cancers Yang, D., Huang, Y., Li, B., Cai, J., Ren, G. 2023; 15 (24)

  Abstract

  In this study, we present an innovative approach that harnesses deep neural networks to simulate respiratory lung motion and extract local functional information from single-phase chest X-rays, thus providing valuable auxiliary data for early diagnosis of lung cancer. A novel radiograph motion simulation (RMS) network was developed by combining a U-Net and a long short-term memory (LSTM) network for image generation and sequential prediction. By utilizing a spatial transformer network to deform input images, our proposed network ensures accurate image generation. We conducted both qualitative and quantitative assessments to evaluate the effectiveness and accuracy of our proposed network. The simulated respiratory motion closely aligns with pulmonary biomechanics and reveals enhanced details of pulmonary diseases. The proposed network demonstrates precise prediction of respiratory motion in the test cases, achieving remarkable average Dice scores exceeding 0.96 across all phases. The maximum variation in lung length prediction was observed during the end-exhale phase, with average deviation of 4.76 mm (±6.64) for the left lung and 4.77 mm (±7.00) for the right lung. This research validates the feasibility of generating patient-specific respiratory motion profiles from single-phase chest radiographs.

  View details for DOI 10.3390/cancers15245768

  View details for PubMedID 38136313

  View details for PubMedCentralID PMC10741831

• Deep learning attention-guided radiomics for COVID-19 chest radiograph classification. Quantitative imaging in medicine and surgery Yang, D., Ren, G., Ni, R., Huang, Y. H., Lam, N. F., Sun, H., Wan, S. B., Wong, M. F., Chan, K. K., Tsang, H. C., Xu, L., Wu, T. C., Kong, F. S., Wáng, Y. X., Qin, J., Chan, L. W., Ying, M., Cai, J. 2023; 13 (2): 572-584

  Abstract

  Accurate assessment of coronavirus disease 2019 (COVID-19) lung involvement through chest radiograph plays an important role in effective management of the infection. This study aims to develop a two-step feature merging method to integrate image features from deep learning and radiomics to differentiate COVID-19, non-COVID-19 pneumonia and normal chest radiographs (CXR).In this study, a deformable convolutional neural network (deformable CNN) was developed and used as a feature extractor to obtain 1,024-dimensional deep learning latent representation (DLR) features. Then 1,069-dimensional radiomics features were extracted from the region of interest (ROI) guided by deformable CNN's attention. The two feature sets were concatenated to generate a merged feature set for classification. For comparative experiments, the same process has been applied to the DLR-only feature set for verifying the effectiveness of feature concatenation.Using the merged feature set resulted in an overall average accuracy of 91.0% for three-class classification, representing a statistically significant improvement of 0.6% compared to the DLR-only classification. The recall and precision of classification into the COVID-19 class were 0.926 and 0.976, respectively. The feature merging method was shown to significantly improve the classification performance as compared to using only deep learning features, regardless of choice of classifier (P value <0.0001). Three classes' F1-score were 0.892, 0.890, and 0.950 correspondingly (i.e., normal, non-COVID-19 pneumonia, COVID-19).A two-step COVID-19 classification framework integrating information from both DLR and radiomics features (guided by deep learning attention mechanism) has been developed. The proposed feature merging method has been shown to improve the performance of chest radiograph classification as compared to the case of using only deep learning features.

  View details for DOI 10.21037/qims-22-531

  View details for PubMedID 36819269

  View details for PubMedCentralID PMC9929417

• Volumetric multiphase ventilation imaging based on four-dimensional computed tomography for functional lung avoidance radiotherapy. Medical physics Huang, Y. H., Ren, G., Xiao, H., Yang, D., Kong, F. S., Ho, W. Y., Cai, J. 2022; 49 (11): 7237-7246

  Abstract

  Current computed tomography (CT)-based lung ventilation imaging (CTVI) techniques derive a static ventilation image without temporal information. This research aims to develop a four-dimensional CT (4DCT)-based multiphase dynamic ventilation imaging framework capable of recovering the entire ventilation process throughout the breathing cycle for functional lung avoidance radiotherapy (FLART).A total of 15 free-breathing thoracic 4DCT scans of lung or esophageal cancer patients were collected from the public datasets. The lung region of each phase image was first delineated, and then the mask-free isotropic total variation image registration algorithm was used to derive the deformation vector fields between the end-expiration (EE) phase and other phases. As a surrogate of ventilation, the voxel-wise local expansion ratio of each phase relative to the EE phase was estimated using the parameterized Integrated Jacobian Formulation method in the EE phase coordinate. Lastly, the dynamic ventilation images were generated by warping these phase-specific local expansion distributions with a same geometry into their respective breathing phases. Quantitative analysis, including interphase Spearman correlation coefficients, voxel-wise, and regional-wise expansion/contraction tracking, were performed to indirectly validate the proposed method.The proposed method maintains the physiological meaning of ventilation on each phase and enables to recover the dynamic lung ventilation process. The mean interphase Spearman correlations ranged between 0.23 ± 0.20 and 0.93 ± 0.04 and decreased near the EE phase. Only 26.2% (2.59E + 6 out of 9.89E + 6) of lung voxels exhibited the same expansion/contraction pattern as the global lung. Qualitative and quantitative evaluations of the interphase ventilation distribution difference show that ventilation spatiotemporal heterogeneities generally exist during respiration.In contrast to conventional CTVI metrics, our method enables to extract additional phase-resolved respiration-correlated information and reflects the generally existed ventilation spatiotemporal heterogeneities. Subsequent studies with quantitative phase-by-phase cross-modality evaluations will further explore its potential to deepen our understanding of lung function and respiration mechanics and also to facilitate more accurate implementation of FLART.

  View details for DOI 10.1002/mp.15847

  View details for PubMedID 35841346

• Radiomics-Based Detection of COVID-19 from Chest X-ray Using Interpretable Soft Label-Driven TSK Fuzzy Classifier. Diagnostics (Basel, Switzerland) Zhang, Y., Yang, D., Lam, S., Li, B., Teng, X., Zhang, J., Zhou, T., Ma, Z., Ying, T. M., Cai, J. 2022; 12 (11)

  Abstract

  The COVID-19 pandemic has posed a significant global public health threat with an escalating number of new cases and death toll daily. The early detection of COVID-related CXR abnormality potentially allows the early isolation of suspected cases. Chest X-Ray (CXR) is a fast and highly accessible imaging modality. Recently, a number of CXR-based AI models have been developed for the automated detection of COVID-19. However, most existing models are difficult to interpret due to the use of incomprehensible deep features in their models. Confronted with this, we developed an interpretable TSK fuzzy system in this study for COVID-19 detection using radiomics features extracted from CXR images. There are two main contributions. (1) When TSK fuzzy systems are applied to classification tasks, the commonly used binary label matrix of training samples is transformed into a soft one in order to learn a more discriminant transformation matrix and hence improve classification accuracy. (2) Based on the assumption that the samples in the same class should be kept as close as possible when they are transformed into the label space, the compactness class graph is introduced to avoid overfitting caused by label matrix relaxation. Our proposed model for a multi-categorical classification task (COVID-19 vs. No-Findings vs. Pneumonia) was evaluated using 600 CXR images from publicly available datasets and compared against five state-of-the-art AI models in aspects of classification accuracy. Experimental findings showed that our model achieved classification accuracy of over 83%, which is better than the state-of-the-art models, while maintaining high interpretability.

  View details for DOI 10.3390/diagnostics12112613

  View details for PubMedID 36359456

  View details for PubMedCentralID PMC9689330

• Deep Learning-Based Automatic Assessment of Radiation Dermatitis in Patients With Nasopharyngeal Carcinoma. International journal of radiation oncology, biology, physics Ni, R., Zhou, T., Ren, G., Zhang, Y., Yang, D., Tam, V. C., Leung, W. S., Ge, H., Lee, S. W., Cai, J. 2022; 113 (3): 685-694

  Abstract

  Radiation dermatitis (RD) is a common, unpleasant side effect of patients receiving radiation therapy. In clinical practice, the severity of RD is graded manually through visual inspection, which is labor intensive and often leads to large interrater variations. To overcome these shortcomings, this study aimed to develop an automatic RD assessment based on deep learning (DL) techniques that could efficiently assist the RD severity classification in clinical application.A total of 1205 photographs of the head and neck region were collected from patients with nasopharyngeal carcinoma (NPC) undergoing radiation therapy. The severity of RD in these photographs was graded by 5 qualified assessors based on the Radiation Therapy Oncology Group guidance. An end-to-end RD grading framework was developed by combining a DL-based segmentation network and a DL-based RD severity classifier, which are used for segmenting the neck region from the camera-captured photographs and grading, respectively. U-Net was used for segmentation and another convolutional neural network classifier (DenseNet-121) was applied to RD severity classification. Dice similarity coefficient was used to evaluate the performance of segmentation. Severity classification was evaluated by several metrics, including overall accuracy, precision, recall, and F1 score.Results of segmentation showed that the averaged dice similarity coefficients were 91.2% and 90.8% for front and side view, respectively. For RD severity classification, the overall accuracy of test photographs was 83.0%. Our method accurately classified 90.5% of grade 0, 67.2% of grade 1, 93.8% of grade 2, and 100% of above grade 2 cases. The overall prediction performance was comparable with human assessors. There was no significant difference in accuracy when using manually or automatically segmented regions (P = .683).We have successfully demonstrated a DL-based method for automatic assessment of RD severity in patients with NPC. This method holds great potential for efficient and effective assessing and monitoring of RD in patients with NPC.

  View details for DOI 10.1016/j.ijrobp.2022.03.011

  View details for PubMedID 35304306

• Development and validation of bone-suppressed deep learning classification of COVID-19 presentation in chest radiographs. Quantitative imaging in medicine and surgery Lam, N. F., Sun, H., Song, L., Yang, D., Zhi, S., Ren, G., Chou, P. H., Wan, S. B., Wong, M. F., Chan, K. K., Tsang, H. C., Kong, F. S., Wáng, Y. X., Qin, J., Chan, L. W., Ying, M., Cai, J. 2022; 12 (7): 3917-3931

  Abstract

  Coronavirus disease 2019 (COVID-19) is a pandemic disease. Fast and accurate diagnosis of COVID-19 from chest radiography may enable more efficient allocation of scarce medical resources and hence improved patient outcomes. Deep learning classification of chest radiographs may be a plausible step towards this. We hypothesize that bone suppression of chest radiographs may improve the performance of deep learning classification of COVID-19 phenomena in chest radiographs.Two bone suppression methods (Gusarev et al. and Rajaraman et al.) were implemented. The Gusarev and Rajaraman methods were trained on 217 pairs of normal and bone-suppressed chest radiographs from the X-ray Bone Shadow Suppression dataset (https://www.kaggle.com/hmchuong/xray-bone-shadow-supression). Two classifier methods with different network architectures were implemented. Binary classifier models were trained on the public RICORD-1c and RSNA Pneumonia Challenge datasets. An external test dataset was created retrospectively from a set of 320 COVID-19 positive patients from Queen Elizabeth Hospital (Hong Kong, China) and a set of 518 non-COVID-19 patients from Pamela Youde Nethersole Eastern Hospital (Hong Kong, China), and used to evaluate the effect of bone suppression on classifier performance. Classification performance, quantified by sensitivity, specificity, negative predictive value (NPV), accuracy and area under the receiver operating curve (AUC), for non-suppressed radiographs was compared to that for bone suppressed radiographs. Some of the pre-trained models used in this study are published at (https://github.com/danielnflam).Bone suppression of external test data was found to significantly (P<0.05) improve AUC for one classifier architecture [from 0.698 (non-suppressed) to 0.732 (Rajaraman-suppressed)]. For the other classifier architecture, suppression did not significantly (P>0.05) improve or worsen classifier performance.Rajaraman suppression significantly improved classification performance in one classification architecture, and did not significantly worsen classifier performance in the other classifier architecture. This research could be extended to explore the impact of bone suppression on classification of different lung pathologies, and the effect of other image enhancement techniques on classifier performance.

  View details for DOI 10.21037/qims-21-791

  View details for PubMedID 35782269

  View details for PubMedCentralID PMC9246721

• Deep learning-based bone suppression in chest radiographs using CT-derived features: a feasibility study. Quantitative imaging in medicine and surgery Ren, G., Xiao, H., Lam, S. K., Yang, D., Li, T., Teng, X., Qin, J., Cai, J. 2021; 11 (12): 4807-4819

  Abstract

  Bone suppression of chest X-ray holds the potential to improve the accuracy of target localization in image-guided radiation therapy (IGRT). However, the training dataset for bone suppression is limited because of the scarcity of bone-free radiographs. This study aims to develop a deep learning-based bone suppression method using CT-derived features to reduce the reliance on the bone-free dataset.In this study, 59 high-resolution lung CT scans were processed to generate the lung digital radiographs (DRs), bone DRs, and bone-free DRs, for the training and internal validation of the proposed cascade convolutional neural network (CCNN). A three-stage image processing framework (CT segmentation, DR simulation, and feature expansion) was developed to expand simulated lung DRs with different weightings of bone intensity. The CCNN consists of a bone detection network and a bone suppression network. In external validation, the trained CCNN was evaluated using 30 chest radiographs. The synthesized bone-suppressed radiographs were compared with the bone-suppressed reference in terms of peak signal-to-noise ratio (PSNR), mean absolute error (MAE), structural similarity index measure (SSIM), and Spearman's correlation coefficient. Furthermore, the effectiveness of the proposed feature expansion method and CCNN model were assessed via the ablation experiment and replacement experiment, respectively.Evaluation on real chest radiographs showed that the bone-suppressed chest radiographs closely matched with the bone-suppressed reference, achieving an accuracy of MAE =0.0087±0.0030, SSIM =0.8458±0.0317, correlation of 0.9554±0.0170, and PNSR of 20.86±1.60. After removing the feature expansion from the CCNN model, the performance decreased in terms of MAE (0.0294±0.0093, -237.9%), SSIM (0.7747±0.0.0416, -8.4%), correlation (0.8772±0.0271, -8.2%), and PSNR (15.53±1.42, -25.5%) metrics.We successfully demonstrated a novel deep learning-based bone suppression method using CT-derived features to reduce the reliance on the bone-free dataset. Implementation of the feature expansion procedures resulted in a remarkable reinforcement of the model performance. For the application of target localization in IGRT, the clinical testing of the proposed method in the context of radiation therapy is a necessary procedure to move from theory into practice.

  View details for DOI 10.21037/qims-20-1230

  View details for PubMedID 34888191

  View details for PubMedCentralID PMC8611463