Head and neck (H&N) cancer is a broad category of diverse cancer types, originating from various soft tissue, glands, and bones (Pai and Westra, 2009). To kill cancerous cells while avoiding normal tissue damage, external radiation therapy (RT) is regarded as the preferred treatment, aiming to deliver a high radiation dose (i.e., prescription dose) to the planning target volume (PTV) while minimizing the dose to organs-at-risk (OARs) via multiple focused radiation beams (Khan, 2010). The radiation delivery of RT is followed by RT plan. Currently, one of the common treatments for external RT is intensity-modulated radiation therapy (IMRT), in which delivered beams are highly conformal to the PTV, and the radiation intensity for each beam can be modulated individually (Webb, 2003). Thus, the RT plan in IMRT is more acceptable physically in the clinical workflow.

Dose distribution design in RT plan is a complex process, including CT image acquisition, ROI contouring on the acquired CT image by radiation oncologist manually (Fig. 1(b) and (e)), treatment parameter arrangement (e.g., geometry of beams), and plan parameter optimization (e.g., dose-volume objectives). The result is a spatial dose distribution, which is called a dose distribution map. The intensity value in each voxel represents the amount of radiation dose accepted by the body in the unit of Gray (Gy) (Khan, 2010).

To obtain clinically acceptable RT plan, dosimetrists need to manually adjust treatment parameters in a trial-and-error manner, this process costs hours so that delaying the best treatment period for each patient (Kearney et al., 2018b). In addition, the quality of RT plan has high variability between inter- and intra-institutions due to differences in technological parameters (including treatment planning system and modality) as well as planner’s skill level (e.g., years of experience and education) (Nelms et al., 2012). The above-mentioned reasons would result in sub-optimal RT plan and thus affect the result of the plan (Peters et al., 2010). Considering the challenge and time-consuming nature of manual RT plan, developing automated methods (e.g., knowledge-based planning Momin et al., 2021, Shiraishi and Moore, 2016, Ge and Wu, 2019, Babier et al., 2020) is of great clinical value.

Recently, due to the fast development of deep learning (DL), especially the convolutional neural network (CNN) as well as its variants (Ronneberger et al., 2015, Milletari et al., 2016), great success has been achieved in solving a broad array of computer vision problems (Liu et al., 2019, Shan et al., 2021, Luan et al., 2008, Jia et al., 2012, Zacharaki et al., 2008). To automatically predict dose distribution map, many DL-based methods have been proposed, which can be generally classified into three categories: 1) designing advanced network architectures, such as C3D (Liu et al., 2021), DoseNet (Kearney et al., 2018a), HD U-net (Nguyen et al., 2019) and DCNN (Gronberg et al., 2021); 2) introducing additional prior knowledge, such as distance map (Zhang et al., 2019) and gradient map (Tan et al., 2021); 3) proposing domain-specific loss functions, such as dose-volume histogram (DVH) loss (Nguyen et al., 2020). However, these DL-based methods are still facing limitations. For instance, the first group of methods often focuses on improving the model’s global performance while losing accuracy for some local regions related to hard-to-learn features. The second group of methods considers physical prior of dose distribution to facilitate the learning of discriminative features while ignoring the geometric prior of beam-shaped radiation in RT, thus causing poor performance along beam paths. The last group of methods utilizes elaborated loss functions to regularize the key indices while imposing high computational overhead and GPU memory consumption on the model training.

In this work, to deal with the aforementioned limitations and achieve high-performance automatic RT plan, we propose a beam-wise dose composition learning (BDCL) method to gradually estimate the dose distribution map in a three-stage (global-beam-global) manner. Specifically, we first employ a global dose network (GDN) to predict a coarse dose distribution over the whole-image space. Then, the coarse dose distribution is decomposed into a series of field doses (beam voters) and further refined by a beam dose network (BDN) according to the geometric prior of the radiation beams. Finally, all the refined beam voters are reassembled into a new global dose distribution, which is further refined by our proposed edge enhancement and DVH calibration processes to meet clinical criteria. We conduct extensive experiments on a public H&N cancer RT dataset, the experimental results show that our method outperforms other state-of-the-art methods by a significant margin and the predicted dose distribution is much closer to the physically deliverable one by using the machine parameters and beam fluence that deliver it. In summary, our main contributions are four-fold:

We propose to generate beam masks as the prior knowledge of beam-wise radiation delivery by a novel beam mask generator, which guide and decompose the coarse dose distribution map into multiple field doses. This process exploits dose distribution on the beam paths in a beam-wise way, which decomposes the difficult task into a few easy-to-learn sub-tasks.

We propose an overlap consistency module to make the predictions of overlapped regions between different beam voters consistent, which improves the accuracy of the prediction and accelerates the convergence speed of the model.

We present a novel multi-beam voting mechanism to reassemble the global dose distribution map from the multiple beam voters, which lays the foundation for global-wise dose refinement.

We integrate DVH metrics into DL model training by the proposed DVH calibration process, which makes the prediction in ROIs more accurate and efficient. Besides, we also apply edge enhancement to enhance boundary learning, making the prediction sharper.

This work is a substantial extension of our conference paper published on MICCAI 2022 (Wang et al., 2022) in the following highlighted aspects. First, we further improve the performance of our method by proposing the overlap consistency module and the edge enhancement process. Second, we conduct comprehensive ablation studies on the proposed method to justify our designs in a more systematic way. Third, we introduce more radiotherapy-specific DVH metrics to evaluate experimental results, demonstrating that our prediction is closer to the real clinical RT plan than predictions of the state-of-the-art methods in terms of clinical criteria. Last, we have thorough discussions on this study, regarding experimental results, strengths, and limitations of the proposed method.