Automated adaptive radiotherapy system with machine learning dose prediction

Abstract

An automated adaptive radiotherapy system based on machine learning for designing and adapting personalized dosing plans and a corresponding system, the system comprising: • a data acquisition module to collect multimodal patient data such as anatomical images, genomics, physiological signals and electronic health records; • a preprocessing department that performs the normalization, alignment, segmentation and transformation of the acquired data into data structures suitable for predictive modeling; • a dose prediction engine containing at least one machine learning model trained to predict personalized three-dimensional radiotherapy dose distributions and dose-volume histograms from the preprocessed data; • an adaptation module to receive updated clinical data and automatically adapt the dosing schedule to intra-fraction and inter-fraction changes in patient anatomy and response to treatment; • a clinician dashboard that displays predicted dosing schedules, allows clinician interaction, and can make the model interpretable through explainable AI; • a data security and compliance layer that protects data through encryption, access control, and regulatory compliance.

Description

Field of the invention

[0001] The present invention relates generally to adaptive radiotherapy systems and, more particularly, to the use of machine learning to predict and optimize dose distributions for radiotherapy based on a patient's anatomical and genomic data. Background of the invention

[0002] Radiotherapy is a cornerstone of cancer treatment and is used in over 50% of all cancer patients as a mono- or multimodal therapy comprising surgery, chemotherapy, and immunotherapy. The goal of radiotherapy is to deliver a cytotoxic dose of ionizing radiation to tumor cells while sparing adjacent tissues and organs at risk (OAR). Traditional radiotherapy procedures are based on a static treatment planning structure, which is based on a simulation session with diagnostic imaging (usually computed tomography (CT), magnetic resonance imaging (MRI), or positron emission tomography (PET) for improved soft tissue delineation and acquisition of metabolic data). This initial plan, the "planning CT," serves as the basis for establishing the contours of the total tumor volume (TTV) and the OAR.These are imported into a treatment planning system (TPS) and used to define dose distributions, which are typically calculated using pencil beam, collapsed cone convolution, or Monte Carlo algorithms.

[0003] However, during the course of fractionated therapy, which extends over many weeks and typically comprises 20–40 fractions, not only the tumor but also normal tissue can undergo significant changes over time due to radiobiological reactions, tumor shrinkage, weight loss, fluid shifts, or organ movement. These changes can lead to inadequate dose coverage of the tumor and inadvertent irradiation of the OAR. For example, a shrunken tumor could result in an overdose in which too much radiation inadvertently impinges on a displaced organ (e.g., bowel or lung), potentially causing injury. Such inter- and intrafractional anatomical variations represent a source of uncertainty and can result in suboptimal dosimetric accuracy of the original plan.

[0004] Adaptive radiotherapy (ART) was developed to eliminate these uncertainties and enable treatment plan adjustments. ART requires reacquiring current imaging data, readjusting target and OAR contouring, optimizing dose distribution, and validating the new plan for clinical treatment. However, existing ART techniques are predominantly manual and require significant labor, including multiple rounds of segmentation, dose recalculations, and physician peer review. This manual method is time-consuming and relies on subjective clinician judgment, leading to interobserver variability and low reproducibility. Furthermore, differences between imaging and treatment timing can be so significant that same-day planning is not feasible.This compromises the clinical feasibility of ART in rapidly developing cancers such as head and neck, lung or gastrointestinal tumors.

[0005] The rise of machine learning (ML) and artificial intelligence (AI) offers us the opportunity to transform ART into a scalable, intelligent process. Convolutional neural networks (CNNs), transformer models, and ensemble learning have achieved high performance in automating image segmentation, organ delineation, and dose prediction. Deep learning has been shown to be applicable to multimodal datasets (imaging, dosimetric maps, clinical parameters, and even radiogenomic feature data) to achieve high accuracy in predicting patient-specific dose distributions, treatment outcomes, and toxicity risks in various studies. Furthermore, ML algorithms can detect hidden patterns in the clinical dataset that may relate to biological responses or recurrence risk prediction, thus opening another avenue to precision in cancer treatment.

[0006] Nevertheless, the implementation of ML in radiotherapy lags behind, and most applications are limited to static planning aids or offline analytics. A key challenge remains to apply ML to a real-time, adaptive radiotherapy system that offers the possibility of seamless data integration, for example, from daily CBCT scans, deformable image registrations, updates from the electronic medical record, and patient-specific genomic information, so that new dose plans can be delivered without more than minimal user input. Any such system should be able to process multimodal, high-dimensional input data, generate clinically interpretable dose predictions, work within the strict time constraints of a radiotherapy workflow, and comply with medical privacy laws such as HIPAA, GDPR, and their regional variants.

[0007] The present invention addresses these unmet needs by providing an automated adaptive radiotherapy system with machine learning dose prediction. We propose a modular infrastructure that combines state-of-the-art ML algorithms with clinical imaging systems, secure data storage, and cloud or edge computing infrastructure. The system leverages data streams from various sources such as volumetric imaging modalities (CBCT, MRI), structured clinical records, previous dose distributions, and optionally information on genomic signatures, e.g., TP53 mutation status and radiosensitivity indices, to build a comprehensive patient model. For complex cases, the VMAT atlas-based approach uses the optimal voxel-level dose distribution estimated from anatomical changes, biological variability, and historical treatment response trends.These are derived from large, annotated datasets of previously treated patients and are based on models trained using supervised learning techniques. The architecture enables online learning, allowing the model to learn its parameters from treatment outcomes and physician feedback.

[0008] In addition, the framework includes automated segmentation pipelines, DVF calculations for dose accumulation, and uncertainty quantification to prioritize potentially ambiguous predictions for manual review. The prediction engine is coupled with a treatment planning system that automatically generates radiation flux maps based on inverse optimization, taking into account dose tolerance values and protocol constraints. A secure physician interface enables plan review with predicted plans and dose-volume histograms. Recommendations can be approved or overridden by the physician. The system is compatible with common DICOM RT formats and can be deployed in local or compatible cloud environments with support for end-to-end encryption.

[0009] This invention revolutionizes radiation therapy by transforming ART from a reactive, clinically guided approach into a proactive, data-driven, and automated process that significantly improves the precision, safety, and efficiency of radiation treatments. It provides oncologists with true real-time decision support and enables the scalability of personalized treatment plans based on both anatomical dynamics and molecular signatures. This is a groundbreaking breakthrough in the era of precision radiation oncology. Summary of the invention

[0010] The invention describes a novel automated adaptive radiotherapy system (AARS) that utilizes innovative machine learning techniques and multimodal clinical data integration for automated real-time plan adjustment to create personalized and clinically safe radiotherapy dosage plans. Unlike current RT strategies that follow static treatment protocols, the present invention establishes a dynamic, intelligent method that continuously learns and adapts the therapy regimen during treatment to optimize clinical efficacy while reducing the radiation dose to healthy tissue.

[0011] At the core of the model is a comprehensive design with six main components. The first component is the data acquisition module, which acts as an intermediary between the system and its clinical data sources. It receives various patient information, such as high-resolution images (CT, MRI, PET, 4D), genomic / molecular features that can influence radiosensitivity, transient physiological signals (respiratory movement / heart rate), and electronic medical records (demographic data, prior treatments, laboratory values, cancer stage). The module is designed for integration and communicates data via standards-compliant APIs, DICOM protocols, HL7 interfaces, and secure cloud repositories.

[0012] The second part (the preprocessing unit) performs essential data empowerment operations. This tool normalizes and arranges input data from diverse sources to establish a consistent connection between imaging techniques and genomic formats. Sophisticated algorithms remove tumors and organs at risk using AI tools such as U-Net for image partitioning and apply deformable image registration methods to monitor changes in body parts over time. The preprocessing pipeline also includes noise reduction, intensity standardization and resolution harmonization, and image preprocessing for machine learning inference.

[0013] The Dose Prediction Engine is the prediction engine. It leverages an ensemble of machine learning models, from CNNs trained on dose distribution patterns to ensemble learning for model robustness and an active learning agent (i.e., reinforcement learning), to generate the decision sequence that optimizes patient care based on dynamic patient data at the point of care. Such models can be trained on large, diverse datasets spanning cancer types, stages, and patient demographics. Radiobiological factors such as dose-volume histogram (DVH) constraints, tumor control probability (TCP), and normal tissue complication probability (NTCP) are integrated into the engine to generate optimal and safe treatment plans.

[0014] Continuous monitoring of treatment efficacy through the Adaptation Module automatically adjusts the treatment plan to remain effective even in the face of patient anatomical or biological changes. This module enables real-time adjustment within and between fractions. It can reoptimize beam geometries, recalculate dose distributions, and resegment structures almost instantly. The system also generates predictive warnings through deviation detection and uncertainty estimation, allowing for timely detection and treatment of unexpected patient reactions.

[0015] For the clinician interface, the system offers a physician dashboard and a clinical interface with AI-generated plans, as well as visualization and interpretation tools. Physicians have various options for viewing dose overlays on 3D anatomical models, comparing different planning scenarios, and assessing the reliability of model predictions. The dashboard integrates decision support features such as evidence-based treatment recommendations, toxicity predictions, and integration with multidisciplinary tumor boards. Importantly, physicians retain full control, as final decisions can be manually overridden to adapt to expert experience and patient preferences.

[0016] To meet strict regulatory requirements and protect sensitive medical data, a data security and compliance framework was integrated into the system. This includes features such as end-to-end encryption for data in transit and at rest, role-based access control (RBAC), continuous audit logging, and secure user authentication. The framework also aims to meet international standards, including the U.S. Health Insurance Portability and Accountability Act (HIPAA), the European Union's General Data Protection Regulation (GDPR), and the International Organization for Standardization (ISO) and International Electrotechnical Commission (IEC) 27001 information security standard.To ensure the privacy of PingAn Group's patients, their data is anonymized and all transactions are recorded in a fully transparent manner to enable traceability and accountability.

[0017] Introduction: The Automated Adaptive Radiotherapy System (AARS) is a potentially groundbreaking innovation in cancer treatment, enabling hyper-personalized, data-driven, and dynamically adaptive radiotherapy. By integrating artificial intelligence, clinical data, and physician monitoring into an intuitive, user-friendly platform, the invention offers greater treatment effectiveness, shorter planning times, enhanced patient safety, and global scalability. Whether implemented in an established or developing hospital, this system offers a powerful, intelligent approach that fits the future of precision oncology. Detailed description of the invention

[0018] Fig. : Block diagram of the automated adaptive radiotherapy system with machine learning-based dose prediction. Fig. shows a block diagram illustrating the main components of the system and the relationships between them, namely: data acquisition module (102), preprocessing unit (104), dose prediction engine (106), adaptation module (108), physician dashboard interface (110), data security and compliance layer (112), cloud integration interface (114), and hardware acceleration layer (116). Detailed description of the invention

[0019] Fig.shows a block diagram of an automated adaptive radiotherapy system with machine learning dose prediction. The system, commonly referred to as System 100, consists of a multitude of interacting hardware and software elements configured to jointly acquire, preprocess, analyze, and adaptively optimize radiotherapy plans from high-dimensional, multimodal patient data. The system is designed for clinical use, for example, in oncology departments, radiotherapy rooms, and cloud-based diagnostic networks. The following is a brief description of the elements of System 100: The data acquisition module (102) is used to acquire high-resolution, patient-specific clinical data from internal and external sources. This data is based on anatomical imaging techniques such as CT, MRI, PET, and 4D CT, which provide comprehensive information on the structure and spatial dynamics of a tumor. Functional imaging, including perfusion- or diffusion-weighted MRI, can also be integrated for metabolic and cellular profiling. The method also enables the upload of genomic data such as WGS, exome sequencing (WES), and RNA sequencing, which can be accessed directly from sequencing devices or cloud-based bioinformatics resources. The module also receives physiological signals such as respiratory cycles, heart rate variability, oxygen saturation, and EMG (electromyography) from wearable biosensors or intensive care monitors.To promote clinical interoperability, it is based on industry standards such as HL7, DICOM, and FHIR, enabling secure data exchange with EHRs, PACS, and LIS solutions. Optional second or additional auxiliary input ports for third-party devices. A preprocessing unit (104) that converts raw clinical data into a structured, machine-interpretable format suitable for further analysis. This module consists of several specialized pipelines: (i) an imaging engine for noise reduction, artifact correction, and spatial normalization using deep learning and conventional filters (e.g.,Gaussian smoothing, non-local averages); (ii) a deformable image registration system for registering sequential and multimodal scans; (iii) an AI-based segmentation subsystem that uses models such as 3D U-Net and V-Net to automatically delineate tumors and OARs; (iv) a genomics pipeline that filters low-quality reads, aligns sequences to reference genomes, identifies mutations, and normalizes data using TPM, Z-score, or log-fold-change approaches; and (v) a physiological signal processor that extracts informative features from time-series data using Fourier transforms, wavelet decomposition, or LSTM-based deep learning models. The final product is a transformed, multimodal feature vector of the patient's spatial, molecular, and physiological status across time points. A dose prediction engine (106) at the intelligent core of the system creates patient-specific radiation dose distributions. The engine uses various machine learning models, including deep convolutional neural networks (CNNs), attention models, graph neural networks (GNNs), and boosted decision trees, trained on large, annotated radiotherapy datasets. The models are trained using spatial dose patterns of three-dimensional voxelized anatomy and account for the heterogeneity of tumor biology and normal tissue response. Radiobiological aspects are addressed through models such as the linear-quadratic model (LQ), the tumor control probability (TCP), and the normal tissue complication probability (NTCP) for biologically based treatment planning.The engine creates dose-volume histograms (DVHs), isodose contours, and optimized beam geometries for photon and proton treatment plans, and also simulates multiple fractionation schemes. An adaptation module (108) dynamically updates treatment plans based on patient data during treatment and between treatment fractions. This module continuously tracks anatomical variations, tumor responses, and physiological trends by easily comparing patient data with expected data at any time. If deviations exceed predefined thresholds (e.g., tumor displacement > 5 mm or unintentional dose deviations > 10% to the OAR), the system initiates automatic replanning. An integrated optimization engine optimizes dose distribution using multi-criteria optimization algorithms to optimally balance tumor coverage and toxicity. Other features include tracking dose accumulation, modeling treatment-related changes (e.g., edema, necrosis), and tumor reduction trend estimation.The module can optionally be connected to airway control or real-time motion tracking devices to enable adaptive therapy in real time. A physician dashboard interface (110) was developed for improved clinical clarity and usability. The dashboard provides an interactive 3D anatomy viewer with overlay analysis tools for predicted and delivered dose distributions, segmented anatomical structures, and visual indications of plan deviations. The control panel allows the user to manually modify contours, view DVH metrics, compare treatment plans, and quantitatively review parameters such as conformity index, homogeneity index, and dose constraints. The user interface incorporates Explainable Artificial Intelligence (XAI) capabilities such as SHAP (SHapley Additive ExPlanations), LIME (Local Interpretable Model-agnostic Explanations), and Grad-CAM to justify the recommendations made by the model. The dashboard is available locally on touchscreen consoles and via web-based and encrypted clients over the internet.Compliance - complete audit trails and manual override features ensure compliance and accountability in clinical decisions. A data security and compliance layer (112) ensures comprehensive protection of sensitive patient data. This layer includes AES-256 encryption for data at rest, TLS 1.3 for data in transit, and strong identity authentication methods such as multi-factor authentication, biometric login, and OAuth-based token validation. System activity, model usage, plan creation, data access, and user interaction are continuously recorded in an immutable blockchain audit trail, ensuring easy traceability. The platform is also GDPR and ISO / IEC compliant, ensuring international healthcare data protection. Optional pseudonymization modules enable secure model training and federated models without compromising private patient data. A cloud integration interface (114) connects the system to external genomic databases, national cancer registries, clinical trial databases such as ClinicalTrials.gov, and multicenter collaborative platforms. The interface enables real-time information exchange for updating and optimizing the federated model, benchmarking population datasets, and retrieving clinical trial results based on patient-specific mutations. Transfer learning modules can be added to enable continuous data training and model quality assurance. A hardware acceleration layer (116) provides high-throughput computing performance. This layer includes parallel processing architectures such as GPUs, TPUs, and FPGAs, which enhance image segmentation, dense model inference, and genomic variant detection. Power management improves computing performance under heavy loads through DVFS.

[0020] The Assembleui framework works modularly but seamlessly together to support clinically driven or autonomous workflows. It supports continuous learning, adaptive planning, and precisely guided treatments—from the most centralized oncology centers to the most remote mobile units and telemedicine-connected clinics—for every cancer type and in every clinical setting. Its extensible architecture and modular integration make it a versatile platform for developing next-generation precision radiotherapy.

Claims

[1] Automated adaptive radiotherapy system based on machine learning for designing and adapting personalized dosing plans and a corresponding system, the system comprising: • a data acquisition module to collect multimodal patient data such as anatomical images, genomics, physiological signals and electronic health records; • a preprocessing department that performs the normalization, alignment, segmentation and transformation of the acquired data into data structures suitable for predictive modeling; • a dose prediction engine containing at least one machine learning model trained to predict personalized three-dimensional radiotherapy dose distributions and dose-volume histograms from the preprocessed data; • an adaptation module to receive updated clinical data and automatically adapt the dosing schedule to intra-fraction and inter-fraction changes in patient anatomy and response to treatment; • a clinician dashboard that displays predicted dosing schedules, allows clinician interaction, and can make the model interpretable through explainable AI; • a data security and compliance layer that protects data through encryption, access control, and regulatory compliance. [2] The system of claim 1, wherein the data acquisition module is further configured to receive image data from CT, MRI, PET, and 4D-CT, as well as physiological data from wearable biosensors. [3] The system of claim 1, wherein the genomic data comprises WGS, WES and RNA sequence data and is harmonized with a standardized genomic preprocessing pipeline. [4] The system of claim 1, wherein the preprocessor comprises a CNN-based image segmentation engine selected from the group consisting of 3DU-Net, V-Net, and DeepLab models. [5] The system of claim 1, wherein the dose prediction engine further comprises a radiobiological modeling unit that implements linear-quadratic models and calculates the tumor control probability (TCP) and the normal tissue complication probability (NTCP). [6] The system of claim 1, wherein the adaptation module comprises a real-time motion prediction unit and the dose recalculation is based on the registration of the deformable images and the motion tracking data. [7] The system of claim 1, wherein the clinician dashboard interface includes a 3D visualization environment to enable interactive contour editing, plan comparison, and real-time visualization of treatment metrics. [8] The system of claim 1, wherein the Explainable Artificial Intelligence (XAI) module includes SHAP, LIME, or Grad-CAM frameworks to highlight input features that influence model decisions. [9] The system of claim 1, wherein the data security and compliance layer comprises: role-based access control (RBAC), multi-factor authentication, and blockchain-based audit logging. [10] The system of claim 1, wherein the functions further comprise a cloud integration interface communicatively coupled to at least one of the following for model refinement: an external genomic database, an external clinical trial repository, and an external federated learning platform. [11] The system of claim 1, further comprising a hardware acceleration layer including at least one GPU, TPU, or FPGA to enable parallel processing of image and genomic data. [12] The system of claim 1, wherein the system is capable of simulating both photon and proton radiotherapy modalities and multiple fractionation schedules. [13] The system of claim 1, wherein the adaptation module is further capable of modeling tumor shrinkage paths and taking into account anatomical variations caused by radiotherapy in treatment planning. [14] The system of claim 1, wherein the system is modular, scalable and transportable in central hospitals, mobile radiotherapy units and remote clinics connected via the cloud.

Images