The incidence of esophageal adenocarcinoma has been increasing over recent decades [1]-[5]. For patients with locally advanced disease, tri-modality therapy—comprising chemotherapy, radiotherapy, and surgical resection—remains the standard of care. Conventional radiotherapy, which employs megavoltage photon beams delivered using three-dimensional conformal radiation therapy (3D-CRT), intensity-modulated radiation therapy (IMRT), or volumetric modulated arc therapy (VMAT), represents the established standard approach in radiotherapy [6].

In the last decade, proton beam therapy (PBT) has emerged as an alternative radiation modality [7]. The defining physical property of proton beams—the Bragg peak—enables precise dose deposition with no exit dose, thereby improving sparing of healthy tissues distal to the target volume. This characteristic represents the principal dosimetric advantage of protons over photons [8]. Comparative studies of PBT versus IMRT/VMAT for esophageal cancer have consistently demonstrated significant reductions in radiation exposure to the heart and lungs [8]-[13]. These findings underscore the potential of PBT to improve treatment outcomes for tumors located near critical organs-at-risk (OARs), including the lungs, heart, kidneys, bowel, stomach, and spinal canal.

As with IMRT, intensity-modulated proton therapy (IMPT) planning relies on an iterative, trial-and-error optimization process that can be time-consuming and subject to inter-planner variability. Enhancing planning efficiency and ensuring consistent plan quality therefore remain major challenges in IMPT.

Knowledge-based planning (KBP) is a data-driven methodology originally developed for IMRT to improve planning efficiency and plan quality consistency [14]-[19]. One KBP strategy developed by our group is the overlap-volume histogram (OVH)-driven KBP method, whose feasibility and effectiveness have been demonstrated across multiple anatomical sites [15]-[19]. Several commercial implementations of KBP are now clinically available [20]-[22].

The present study extends the OVH-driven KBP approach to IMPT for esophageal cancer. Its primary objectives are:

To apply the OVH-KBP methodology, originally developed for IMRT, to IMPT in a cohort of esophageal cancer patients; andTo evaluate whether the KBP-guided IMPT plans are comparable in quality to manually created clinical plans (CPs) while improving planning efficiency. Here, CPs refer to IMPT plans manually generated by planners and clinically delivered to patients.

As a side note, planning and optimization time for KBP generation was not recorded in this study, as these metrics have been reported in our previous publications [15][19].

Figure 1.Average achieved DVHs for the GTV, CTV and OARs for CTV prescription of 41.4 Gy (RBE). Data is shown for CPs (red line) and KBPs (blue line).

Figure 2.Average achieved DVHs for the GTV, CTV and OARs for CTV prescription of 45 Gy (RBE). Data is shown for CPs (red line) and KBPs (blue line).

Figure 1 and Figure 2 present the achieved average DVHs for the CPs and KBPs across target volumes and OARs for prescription levels 41.4 Gy and 45 Gy respectively. Quantitative dose-volume parameters extracted from individual patient DVHs are summarized in Tables 1-4, including mean values, clinical planning goals, and interpatient standard deviations.

Table 1. Summary of the planning objectives of the CTV prescription of 41.4 Gy (RBE).

Table 2. Summary of the planning objectives of the CTV prescription of 45 Gy (RBE).

Table 3.Summary of the average results for the OARs of the CTV prescription of 41.4 Gy (RBE).

Table 4.Summary of the average results for the OARs of the CTV prescription of 45 Gy (RBE).

The aim of this study was to apply a KBP framework for IMPT in patients with locally advanced esophageal cancer. The framework generated feasible DVH objectives using a database containing geometric and dosimetric information from previously treated patients. The results demonstrate that the proposed KBP model can efficiently generate IMPT plans with plan quality comparable to that of clinically implemented, manually generated plans by human planners.

This work employed proton beams delivered by a Mevion S250i proton therapy system, which utilizes an energy modulation mechanism distinct from those of other commercial systems [24]. The Mevion system produces proton energies through the binary insertion of polycarbonate plates of varying thicknesses into the beam path, whereas most other systems use magnetic dipoles or quadrupoles for energy selection. Consequently, the Mevion S250i generally produces larger spot sizes but achieves substantially shorter energy-switching times. To compensate for the dosimetric effects of larger spot sizes, a spot-size-trimming device, the Adaptive Aperture (AA), is employed to sharpen the beam penumbra. These unique beam characteristics directly influence the dose distribution and differentiate the Mevion system from other proton platforms. Therefore, the dosimetric outcomes reported here may not be directly transferrable to other proton therapy systems. Extending this KBP methodology to other systems would require the development of a clinical plan database specific to proton platforms of comparable beam design.

The database utilized in this study was relatively small and constructed from plans generated at a single institution, which inevitably introduces some limitations in terms of plan diversity and consistency. Expanding this database to include plans from multiple institutions—particularly those using the same proton therapy system (e.g., Mevion S250i utilized in this study)—would enhance model robustness and facilitate the generation of more consistent, high-quality plans across diverse clinical environments. Such an expanded dataset would also improve planning efficiency and reduce variability between planners.

Anatomical characteristics of the distal esophagus and adjacent structures influenced beam arrangement decisions, which typically consisted of right and left posterior oblique beams at approximately 140˚ and 220˚. Previous investigations in liver treatment have indicated that minor variations in beam angles exert minimal influence on plan quality [14]. However, larger deviations from the modeled gantry angles could lead to suboptimal optimization outcomes because robust optimization is highly dependent on the geometric relationships between beam paths, targets, and OARs [25]. If different esophageal segments were considered, alternative beam configurations—such as anterior oblique arrangements—may be required. This further emphasizes the need for a larger and more diverse database encompassing the full range of anatomical and geometric scenarios encountered in clinical practice.

All treatment plans in this study were normalized to the mean dose to the CTV, although other institutions may adopt different normalization practices. One of the central aims of KBP is to standardize the planning process, thereby reducing variability introduced by individual planner preferences and experience. Expanding the database to encompass a broader range of target volumes and geometric relationships between targets and OARs will strengthen the KBP model’s ability to produce consistent, high-quality plans across varying patient anatomies. Another limitation of this study stems from the wide variation in esophageal length among patients; in some cases, two isocenters and two fields per isocenter were required to achieve full target coverage, rather than a single isocenter with two fields.

The favorable dosimetric outcomes demonstrated in this study, together with the strong performance of our KBP framework, underscore the value of our patient database as a foundational resource for institutions aiming to implement PBT for advanced esophageal cancer. Notably, the average spinal canal doses—D_0.1cc and D_1%—in the KBP-generated plans were approximately 8 Gy lower than those in the CPs, suggesting a meaningful reduction in the risk of radiation-induced myelopathy if KBP is integrated into IMPT planning. In addition, the KBP plans achieved significantly lower mean dose and D_1% for both the GTV and CTV. A lower D_1% indicates reduced hotspot regions within the targets, which can be advantageous when a more homogeneous dose distribution is desired. However, the broader applicability of our approach across institutions may be affected by contouring heterogeneity among physicians and institutions. Because the OVH—a key geometric descriptor in KBP—depends on the accuracy and consistency of contour delineations, variations in contouring practices could influence model performance. The next phase of this research will evaluate the KBP model using datasets from multiple institutions to assess the impact of contouring variability on plan quality and generalizability. This systematic investigation aims to elucidate the extent to which inter-institutional differences in contouring influence the efficacy of KBP models and to inform strategies for harmonizing contouring standards across centers.

This study demonstrates that the KBP methodology originally developed for IMRT can be effectively adapted to IMPT. By leveraging geometric and dosimetric data from previously delivered treatment plans, this approach enables an optimal balance between target coverage and OAR sparing. The KBP framework provides a reliable and consistent method for predicting clinically achievable DVH objectives prior to optimization, thereby facilitating the generation of high-quality treatment plans with minimal manual adjustment. These findings support the clinical feasibility of KBP-based IMPT planning and underscore its potential to enhance planning efficiency and standardization in proton therapy for esophageal cancer. As a next step, we aim to validate the proposed KBP-IMPT strategy across multiple institutions.

Boran Zhou:Performed OVH, DVH calculations, data analysis and presentation in figures and tables; writing of the original draft of manuscript. Binbin Wu: Provided software for OVH calculations, consultation on the methodology. Markus Wells: Compiled patient plans for library data construction. Keith Unger: Reviewed and edited the manuscript. Dalong Pang: Originated and supervised this research; advised on manuscript preparation and reviewed and edited the manuscript.