Artificial intelligence-powered innovations in radiotherapy: boosting efficiency and efficacy

Introduction

Cancer has become a major global health burden, with its incidence rates continuing to rise. By 2050, there will be about 35.3 million new cancer cases worldwide, representing an increase of approximately 77 % compared to 20.0 million cases in 2022 [1]. In cancer treatment, radiotherapy (RT) plays a crucial role as either a curative or palliative treatment, benefiting around 50 % of cancer patients. However, due to challenges such as limited infrastructure, technology, and human resources, millions of patients worldwide are currently unable to access this vital treatment [2].

The rapidly advancing field of artificial intelligence (AI) has demonstrated intelligence on par with, or even surpassing human capabilities. AI has been applied to various processes in RT, enhancing both the efficiency and efficacy of treatments [2], particularly in time-consuming yet crucial tasks like target delineation and treatment planning [3], [4], [5], [6], [7] (Figure 1).

Figure 1:

The radiotherapy (RT) processes, each of which can be enhanced by artificial intelligence (AI) to improve efficiency and outcomes. Initially, during the patient’s first consultation, AI assists in decision-making. Following this, a CT scan of the target area is performed, with AI helping to enhance image quality for better accuracy. Based on CT images, treatment planning is carried out, where AI improves both the efficiency and effectiveness of contouring and RT planning compared to conventional methods. After undergoing quality assurance, the patient proceeds with treatment delivery, where AI aids in image-guided patient positioning to ensure precision. Finally, post-treatment, AI is used for toxicity prediction, helping to inform the patient’s prognosis.

However, current AI in RT shows limited efficacy improvements and rarely surpasses human expertise [3]. Digital twins are technology that simulate and analyze physical objects, processes or systems by constructing virtual models. In the era of precision medicine, radiotherapy digital twins (RDTs) combined with AI technology promote the integration and processing of data [8], making it possible to personalize RT and improve efficacy.

AI-driven target volume contouring

The effectiveness of RT largely depends on the accuracy of target delineation, which must strike a balance between protecting the organs at risk (OARs) and delivering the prescribed dose to the target tumor. Physicians typically contour OARs and target tumors manually. It can take several hours or even days, making it one of the most time-consuming tasks in clinical practice. Furthermore, manual delineation lacks reproducibility. AI-based segmentation methods, particularly those relying on visual models, significantly enhance the efficiency and reproducibility of RT planning [3], reducing the time required for this process to mere minutes.

For example, Feng Shi et al. proposed a lightweight deep learning framework for RT planning, RTP-Net, which can accurately contour OARs within seconds [4]. Meanwhile, Alexander Kirillov et al. introduced the highly versatile Segment Anything model (SAM) [9], which has been shown to effectively contour OARs, although its ability to contour target tumors remains limited.

Previously, models mainly relied on unimodal imaging data, which was of a single type and had inherent limitations. Tumor staging, metastatic dissemination, and other clinical factors such as the patient’s overall health status, significantly impact the accuracy of tumor delineation. Recently, Yujin Oh et al. introduced a multimodal AI named LLMseg, which integrates key clinical information using large language models (LLMs) [5]. This novel approach has consistently outperformed vision-only AI models, especially in scenarios with limited data availability. In some cancer types, it even shows tumor delineation capabilities comparable to those of clinical experts. Additionally, this technology can adapt its delineation strategies based on various clinical data, significantly enhancing the potential for personalized RT. However, further validation across additional tumor types is still needed to confirm its broader applicability.

AI-driven automated dose prediction and RT planning

RT planning is a crucial step in RT, directly impacting the quality of treatment. Physical therapists must constantly refine the plan to meet dose prescription requirements and achieve optimal therapeutic outcomes. This process is not only time-consuming but also requires a delicate balance between treatment effectiveness and operational efficiency. As a result, outcomes may be suboptimal in some cases, or the planning process may become excessively time-intensive.

The integration of AI for dose prediction can relieve physical therapists from these labor-intensive tasks, thereby enhancing both the efficiency and quality of RT planning. Traditional automation methods include plan-based automated iterative optimization (PB-AIO), multi-criteria optimization (MCO), and knowledge-based planning (KBP). Although these methods can improve planning efficiency to a certain degree, they are limited by the lack of spatial dose distribution information. As a result, they fail to fully satisfy the requirements of personalized and precise treatment [10].

Recent breakthroughs in deep learning techniques, like U-Net and its variants, have substantially enhanced dose prediction. These methods automatically extract features from inputs such as CT images, thereby boosting accuracy [7]. Compared to traditional machine learning, deep learning has a remarkable ability to capture global features. This results in greater efficiency and improved precision, especially when dealing with complex treatment plans. Additionally, the multi-task deep neural network named Deep Profiler incorporates radiomics information to achieve personalized dose prediction [6]. It is predictable that leveraging multimodal AI, integrated with LLMs, to combine CT images with patients’ clinical data will drive forward dose prediction. This, in turn, will contribute to the development of more optimized radiation treatment plans.

All-in-One RT

AI technology has significantly enhanced the efficiency of various RT processes, including autosegmentation, autoplanning, image guidance, beam delivery, and in vivo quality assurance (QA). These processes have been streamlined to be completed within seconds or minutes. Lei Yu and colleagues introduced the concept of All-in-One RT, which integrates these procedures into a single protocol, with critical decision points addressed while the patient remains on the treatment couch throughout the entire process [11]. This innovation has reduced the interval for the first RT treatment from several days to approximately 20 min. In the treatment of rectal cancer patients, it has also minimized uncertainties related to patient anatomy and machine calibration. With the integration of more advanced models for automated treatment planning and other technologies, the future will see the development of the All-in-One RT that offers both enhanced efficiency and effectiveness, potentially eliminating the need for human decision-making in the process.

Construction and framework of RDTs

Current AI in RT primarily focuses on improving efficiency, with limited impact on therapeutic efficacy. Integrating RDTs with AI holds promise for enhancing clinical outcomes while maintaining efficiency. Digital twins have demonstrated great potential in the realm of cancer care and treatment [12], [13], [14].

After gathering and pre-processing data related to patients’ tumor conditions, including tumor location, size, and invasion extent, as well as individual characteristics such as omics, medical history and clinical manifestations [8], 12], 15], an RDT of individual patient is constructed in the digital realm.

Considering the mechanism of RT at various scales, the proposed RDTs framework of individual patient can be technically subdivided into two levels. The physical level digital twin focuses on simulating the geometric modeling of organs and tumors, as well as ray absorption. Meanwhile, the biological-level digital twin is in charge of tracking changes in signaling pathways, alterations in the tumor microenvironment, and biomarker responses (Figure 2A). The connection across different dimensions is realized through information sharing and interaction between these two levels.

Figure 2:

Leveraging radiotherapy digital twins (RDTs) in conjunction with artificial intelligence (AI) for precision therapy. (A) Construction and framework of RDTs. (B) A radiotherapy response prediction model developed using RDTs.

Application of RDTs in conjunction with AI

During the RT process for tumor patients, RDTs are constructed based on patient historical data and updated with real-time data. Treatment plans including target delineation and dosing scheme can be initially developed with reference to similar cases found using the nearest neighbor search. Subsequently, the RT prediction model is employed to optimize the treatment plan (Figure 2B), aiming to achieve the most effective RT outcome.

Anirban Chaudhuri and colleagues proposed a predictive digital twin methodology to optimize dosing scheme for high-grade gliomas under uncertainty [14]. The digital twins integrating patient MRI data with a tumor growth model via Bayesian calibration, propose a suite of optimal dose regimens that balances tumor control and toxicity by solving a multi-objective optimization problem.

For target contouring, similar to LLMSeg integrating clinical data to adjust segmentation models through the LLM [5], RDTs comprehensively and deeply integrate a wide range of multi-dimensional clinical data, and are expected to be used better in personalize tumor contouring for patients in the future.

Challenges

In the development of AI tools for RT, several significant hurdles exist. Firstly, the scarcity of high-quality data for training and validating AI models severely undermines the accuracy of these models. Secondly, data collection and processing also present privacy-related challenges [8]. Moreover, in clinical practice, the “black box” nature of AI makes the decision-making process of doctors obscure, resulting in accountability issues and a crisis of trust [2].