Deep learning to improve dose reconstruction for adaptive radiotherapy in real time

Abstract

A computer-implemented system for improving radiation dose reconstruction in adaptive real-time radiotherapy, comprising a data acquisition module for capturing treatment data in real time during radiotherapy, a preprocessing module for preprocessing the captured data, a deep-learning dose reconstruction engine for reconstructing the delivered radiation dose distribution using a trained deep-learning model, a dose evaluation and comparison module for comparing the reconstructed dose distribution with a planned dose distribution, and an adaptive decision support module to assist in adjusting radiotherapy treatment parameters based on the comparison results.

Description

Application area of ​​the invention

[0001] The present invention relates generally to the field of radiotherapy and medical imaging systems. In particular, it relates to computer-aided methods and systems that use deep learning and artificial intelligence to reconstruct the radiation dose during radiotherapy. The invention specifically targets adaptive real-time radiotherapy, in which the reconstructed dose information is used for dynamic treatment monitoring, verification, and adjustment to improve the accuracy and safety of cancer treatment. Background of the invention

[0002] Radiation therapy is a cornerstone of modern cancer treatment and is frequently used either as a primary therapy or in combination with surgery and chemotherapy. In radiation therapy, high-energy ionizing radiation beams are precisely shaped and directed at the malignant tissue to destroy cancer cells while minimizing radiation exposure to surrounding healthy organs and vital structures. The clinical effectiveness of radiation therapy depends significantly on the precise dose delivery in space and time.

[0003] Conventional radiation therapy planning takes place before the actual treatment and includes pre-therapeutic imaging using computed tomography (CT), magnetic resonance imaging (MRI), or positron emission tomography (PET). Based on these images, target volumes and organs at risk (OAR) are contoured, and the radiation parameters for a planned dose distribution are determined. This is optimized under the assumption of stable patient anatomy and position during radiation therapy. However, this assumption is frequently not met in clinical practice for various reasons.

[0004] Patient-related variations, such as those caused by respiration, involuntary movements, anatomical changes, or tumor volume reduction and organ filling, lead to significant spatial deviations between the radiation beams and the target tissue during treatment. Furthermore, system-related uncertainties such as mechanical tolerances of multileaf collimators (MLCs), gantry rotation, dose rate fluctuations, and field transmission errors cause additional discrepancies between the planned and the actually delivered dose.

[0005] To compensate for these differences, dose reconstruction methods are used to derive the applied dose distribution from treatment log files (TLFs), dosimetric measurements, and intratherapeutic images. Modern dose reconstruction practice: Traditional dose reconstruction methods are predominantly based on physically based dose calculation models, such as analytical and Monte Carlo simulations of the absorbed dose. Although MC methods are considered the gold standard for dose accuracy, they are computationally expensive and have long execution times, making them unsuitable for real-time or near-real-time clinical use.

[0006] Many of the dose reconstruction methods available today are therefore retrospective. This offline analysis does not allow physicians to identify and correct dosimetric deviations during the same treatment session and significantly limits their ability to adjust treatment parameters to these dynamic changes. This poses a particular challenge with modern radiotherapy techniques such as IMRT, VMAT, SBRT, or MR-guided radiotherapy, as the dose gradients are steep and even the smallest errors can have a significant impact on clinical success.

[0007] Adaptive radiotherapy has become established as a concept designed to address these challenges and allows the treatment plan to be adapted to patient-specific changes during the course of therapy. Adaptive strategies can include reoptimizing radiation parameters, changing dose limits, or modifying the irradiated volume. However, the effectiveness of adaptive radiotherapy depends significantly on the availability of fast and highly accurate dose reconstruction tools in a mode suitable for use during or immediately after treatment.

[0008] To date, the real-time application of adaptive radiotherapy systems has been limited by the computation time of conventional dose calculation and reconstruction methods. Current adaptive systems use simplified dose models or delayed dosimetry analyses, which can impair accuracy or responsiveness during treatment. Traditional methods also exhibit weaknesses with noisy, incomplete, and / or indirect data—for example, with low-quality portal images or sparsely sampled dosimetry values.

[0009] Advances in artificial intelligence and deep learning in recent years have demonstrated that data-driven models can learn complex, nonlinear mappings between input data and target outputs for a wide range of applications in medical imaging and radiation therapy. Deep neural networks, including CNNs, RNNs, and attention-based models, have proven highly effective in various tasks such as dose prediction, image reconstruction, organ segmentation, and treatment outcome modeling. Once trained, these models can draw conclusions with extremely low latency, making them ideally suited for use in real-time scenarios.

[0010] However, existing deep learning-based approaches in radiotherapy largely focus on dose prediction before treatment rather than real-time dose reconstruction during irradiation. Many current solutions do not adequately account for dynamic patient movements, temporal dose accumulation, or real-time integration into radiotherapy systems. Furthermore, few approaches offer mechanisms for uncertainty estimation or seamless integration into adaptive decision-making processes in radiotherapy.

[0011] Therefore, there is an unmet need for an intelligent dose reconstruction framework that combines the accuracy of physics-based methods with the speed of deep learning inference. Such a framework should be able to process heterogeneous real-time inputs, reconstruct three-dimensional (3D) dose distributions with high accuracy, account for patient-specific and system-related variations, and generate online feedback for adaptive radiotherapy.

[0012] The present invention solves these and other problems by providing a deep learning system and method to improve dose reconstruction during radiotherapy, thus enabling adaptive radiotherapy in real time with increased precision, safety and clinical effectiveness. Summary of the invention

[0013] The present invention relates to a computer-aided system and method for improving adaptive radiation dose reconstruction in real time using deep learning. The invention overcomes the technical limitations of known dose reconstruction methods by precisely and efficiently determining the applied radiation dose at the time of treatment.

[0014] In one aspect of the invention, data is acquired in real time during radiotherapy, including intratherapeutic image data, parameters of the radiotherapy device, and patient-related motion and position information. The acquired data is preprocessed and fed as input to a trained deep learning model. This model is configured to reconstruct a three-dimensional radiation dose distribution that represents the dose actually delivered to the patient.

[0015] In one embodiment, the deep learning model comprises one or more neural network architectures selected from convolutional neural networks (CNNs), recurrent neural networks (RNNs), attention-based networks, transformer models, or hybrid combinations thereof. The model is trained using historical radiotherapy datasets containing reference dose distributions obtained through high-precision dose calculation methods or clinical measurements. After training, the model performs dose reconstruction with low computational latency, making it suitable for real-time or near-real-time clinical workflows.

[0016] In another aspect of the invention, the reconstructed dose distribution is continuously or periodically compared with a planned dose distribution and predefined clinical boundary conditions. If deviations are detected that exceed one or more threshold values, the system generates adaptive feedback to support the adjustment of treatment parameters, including beam configuration, dose rate, irradiation time, or irradiation margins. This enables adaptive radiotherapy in real time or during irradiation without interrupting the treatment.

[0017] In another aspect, the system is adapted to calculate confidence or uncertainty measures for the estimated dose in order to increase clinical reliability and facilitate the treating physician's decision-making. The system can also enable cumulative dose tracking across multiple treatment fractions.

[0018] The present invention offers several technical advantages, including a significant reduction in the computation time for dose reconstruction, higher accuracy in noisy, partial, or indirect measurements, easy integration into adaptive radiotherapy systems, and improved patient safety and treatment effectiveness. The invention is suitable for various radiotherapy techniques, such as intensity-modulated radiotherapy (IMRT), volumetric modulated arc therapy (VMAT), stereotactic radiotherapy, and image- or MR-guided radiotherapy systems. Detailed description of the invention

[0019] Fig.Figure 1 shows a block diagram of a deep learning-based system for improving radiation dose reconstruction in adaptive real-time radiotherapy according to an embodiment of the present invention.The system includes a data acquisition module (100) for capturing treatment data in real time, a preprocessing module (200) for preparing the captured data for analysis, a deep learning dose reconstruction engine (300) for reconstructing the applied radiation dose distribution, a dose assessment and dose comparison module (400) for comparing the reconstructed dose with a planned dose distribution, an adaptive decision support module (500) for supporting treatment adjustment, an optional uncertainty estimation module (600) for estimating confidence measures for the reconstructed dose, and an output and visualization module (700) for displaying the reconstructed dose information and adaptive recommendations for the clinical user.

[0020] The present invention describes a deep learning-based system for improving radiation dose reconstruction in adaptive real-time radiotherapy. The system is designed to accurately estimate the actual radiation dose applied during treatment and to enable timely adjustment of the radiotherapy parameters. It overcomes the technical limitations of conventional dose reconstruction methods by combining heterogeneous real-time treatment data with trained deep learning models, resulting in low latency. The system can be implemented as a dedicated computing platform, as a distributed processing architecture, or as an integrated component of an existing radiotherapy system.

[0021] The system comprises several integrated functional modules that enable real-time dose reconstruction, analysis, and adaptive decision support. Each module can be programmed in hardware, firmware, or a combination of both to execute one or more processes. The modules can communicate via wired or wireless interfaces and process data on one or more processors.

[0022] A data acquisition module (100) is designed to capture treatment data in real time during radiotherapy. The data acquisition module (100) can communicate directly with a radiotherapy device, an imaging device, patient monitoring devices, and control systems. The received data can include intratherapeutic imaging data such as cone-beam computed tomography (CBCT), magnetic resonance imaging (MRI), fluoroscopy images, or images from the electronic portal imaging (EPID) system mounted on the C-arm of a clinical linear accelerator. The data acquisition module (100) can also retrieve the treatment device's log files containing parameters such as beam guidance, gantry angles, collimator settings, multileaf collimator blade positions, beam energies, dose rates, and time information. Additionally, patient-related data such as patient position information and motion data (e.g.,Signals from a motion detection probe or a system integrated into components of the portal imaging system and related to respiratory and / or cardiac movements as well as the status of the immobilization system are acquired. The data acquisition module (100) can continuously, periodically, or event-driven acquire data during a treatment session and transmit the acquired data to downstream modules in real time or near real time.

[0023] The received data is passed to a preprocessing module (200) to prepare the raw data for deep learning. The preprocessing module (200) can perform one or more preprocessing operations, such as smoothing, suppression of artifacts and noise, enhancement of image contrast and detail, spatial or temporal resampling, and alignment of the coordinate systems between image data and treatment data. Additionally, the preprocessing module (200) can perform data normalization, scaling, coding, and feature extraction to provide structured input representations for a deep learning model. In one embodiment, when multiple data sources are involved, the preprocessing module (200) fuses the image data with machine parameters and patient movement information to form a consolidated representation.The preprocessing module (200) generates a normalized, machine-readable input that maintains known clinical spatial and temporal intervals.

[0024] The processed data are then sent to the deep learning dose reconstruction engine (300), which is the central processing unit of the system. The deep learning dose reconstruction engine (300) reconstructs a three-dimensional radiation dose distribution that represents the actual dose delivered to the patient. The engine uses a deep neural network architecture selected from various approaches: convolutional neural networks (CNNs) for extracting spatial features, recurrent neural networks (RNNs) or temporal convolutional networks (TCNs) for modeling time-dependent dose accumulation, attention-based networks or transformer architectures for contextual modeling of treatment dynamics, and hybrid or ensemble architectures of the aforementioned network architectures.The deep learning model can be trained using historical radiotherapy data, which may include treatment data, imaging data, and the corresponding reference dose distributions. These reference dose distributions are calculated using a high-precision dose calculation method or clinical dosimetry measurements. Training can be performed using supervised, semi-supervised, or self-supervised learning. After training, the Deep Learning Dose Reconstruction Engine (300) performs low-latency inference to reconstruct the dose distributions during continuous treatment.

[0025] The reconstructed dose distribution generated by the deep learning dose reconstruction module (300) is sent to a dose evaluation and comparison module (400). This module compares the reconstructed dose distribution with a target dose distribution and predefined clinical requirements. For example, it can calculate dose-volume histograms, determine target volume coverage, assess the dose burden on organs at risk, and determine deviation metrics between the reconstructed and planned dose. Module (400) can also track the cumulative dose across multiple treatments and fractions, detecting systematic or transient deviations. The evaluation results provide quantitative values ​​for dose accuracy and safety.

[0026] An adaptive decision support module (500) provides individualized decision support based on evaluation results. The module can trigger an alarm if dose deviations exceed predefined tolerances or clinical limits. Furthermore, it can suggest changes to treatment parameters by varying beam angles, dose rates, MLC positions, irradiation timing, or treatment margins. In various configurations, the module can be integrated into treatment planning / control systems to support automatic or semi-automatic adjustments, or it can operate in "clinician-in-the-loop" mode, providing decision support without directly controlling the irradiation.

[0027] The system includes, among other things, an uncertainty estimation module (600) that estimates the uncertainty or confidence in the reconstructed dose distribution. This module can utilize various probabilistic deep learning methods, ensemble methods, Bayesian inference, or variance estimation techniques to measure the prediction uncertainty. The uncertainty information can be linked to spatial dose ranges or even global dosimetric quantities to increase clinical plausibility and reliability.

[0028] The device also includes an output and visualization module (700) that provides reconstructed dose information, evaluation analyses, uncertainty displays, and adaptive measures suggested by clinicians. The output and visualization module (700) can visualize 3D dose distributions, differential dose distributions, dose-volume metrics, or alerts in real time. The module can include a user interface and be integrated into existing radiotherapy workstations or control consoles, allowing clinicians to access treatment delivery analysis and system alerts during therapy.

[0029] By simultaneously executing the data acquisition module (100), the preprocessing module (200), the deep learning dose reconstruction engine (300), the dose assessment and comparison module (400), the adaptive decision support module (500), and optionally the uncertainty estimation module (600) and the output and visualization module (700), the applied radiation dose during radiotherapy can be estimated quickly and potentially precisely. By using deep learning inference instead of computationally intensive, physics-based simulations, the system achieves very low latency while maintaining clinically acceptable accuracy. Thus, the system supports adaptive radiotherapy in real time, improves treatment accuracy and patient safety, and contributes to informed clinical decisions in dynamic radiotherapy environments.

Claims

[1] A computer-implemented system for improving radiation dose reconstruction in adaptive real-time radiotherapy, comprising a data acquisition module for capturing treatment data in real time during radiotherapy, a preprocessing module for preprocessing the captured data, a deep learning dose reconstruction engine for reconstructing the delivered radiation dose distribution using a trained deep learning model, a dose evaluation and comparison module for comparing the reconstructed dose distribution with a planned dose distribution, and an adaptive decision support module to assist in adjusting radiotherapy treatment parameters based on the comparison results. [2] System according to claim 1, wherein the data acquisition module is configured to acquire at least one of the following data: imaging data during treatment, protocol files of the treatment machine, parameters of the beam guidance, positions of the multileaf collimator or movement data of the patient. [3] System according to claim 1, wherein the preprocessing module is configured to perform at least one of the following steps: noise reduction, artifact correction, spatial normalization, temporal synchronization, coordinate alignment, feature extraction or data fusion. [4] System according to claim 1, wherein the Deep Learning Dose Reconstruction Engine comprises at least one neural network selected from a Convolutional Neural Network, a Recurrent Neural Network, a Temporal Convolution Network, an Attention-based Network, a Transformer Network or a hybrid combination thereof. [5] System according to claim 1, wherein the Deep Learning Dose Reconstruction Engine is trained using historical radiotherapy datasets comprising treatment data and corresponding reference dose distributions. [6] System according to claim 1, wherein the Deep Learning Dose Reconstruction Engine is configured to reconstruct a three-dimensional radiation dose distribution in real time or near real time during treatment. [7] System according to claim 1, wherein the dose assessment and comparison module is configured to calculate dose-volume metrics for target volumes and organs at risk. [8] System according to claim 1, wherein the dose assessment and comparison module is configured to detect dose deviations that exceed predefined clinical thresholds. [9] System according to claim 1, wherein the adaptive decision support module is configured to generate warnings or notifications when dose deviations exceed predefined tolerance limits. [10] System according to claim 1, wherein the adaptive decision support module is configured to recommend the modification of at least one treatment parameter selected from the beam configuration, dose rate, time of irradiation, position of the multileaf collimator or treatment margin. [11] System according to claim 1, further comprising an uncertainty estimation module configured to estimate uncertainty or confidence measures related to the reconstructed dose distribution. [12] System according to claim 11, wherein the uncertainty estimation module uses probabilistic modeling, ensemble learning, Bayesian inference or variance estimation to quantify the prediction uncertainty. [13] System according to claim 1, further comprising an output and visualization module configured to display reconstructed dose distributions, dose deviation maps, uncertainty indicators or adaptive recommendations for a clinical user. [14] System according to claim 1, wherein the system is integrated with at least one radiotherapy modality selected from intensity-modulated radiotherapy, volumetric modulated arc therapy, stereotactic radiotherapy, image-guided radiotherapy or MRI-guided radiotherapy. [15] A non-volatile, computer-readable medium that stores instructions which, when executed by one or more processors, cause the processors to implement the system according to claim 1.

Images