Radiation therapy (RT) for patients receiving definitive treatment for locally advanced non-small cell lung cancers (NSCLC) has been linked to post-RT cardiac toxicities (1). A number of promising phase II trials reported increased tumor control through radiation dose escalation via stereotactic fractionations (2–4). The Radiation Therapy Oncology Group (RTOG) 0617 phase III dose escalation trial compared 74 Gy with the standard 60 Gy dose with concurrent chemotherapy (5). In this study, the higher prescription dose resulted in worse overall survival. Furthermore, cardiac volumes receiving 5 Gy and 30 Gy were found to be linked with higher death rates. In a multivariate analysis, an increased radiation dose to the heart was independently linked with worse overall survival (6). More recent studies have reported the link between radiation doses to cardiac substructures and major adverse cardiac events, including mortality (1, 7). Atkins et al. reported coronary artery disease progression and the risk of ischemic events due to radiation exposure to arterial vessels, especially the left anterior descending artery (LADA) (7). Skytta et al. reported an increase of serum troponin T>30% as an endpoint to greater than 10% of LADA receiving 15 Gy (8). McWilliam et al. identified the base of the heart as a dose-sensitive region linked with poorer survival in lung cancer patients. In this study, patients receiving a mean dose of more than 8.5 Gy to this region were found to have significantly poorer survival (9). In another study, the maximum dose to the LADA of greater than 53 Gy was linked to increased risks of cardiac morbidity (10). The maximum dose to the right atrium, right coronary artery, and ascending aorta were found to be the most important factors associated with survival (11). Gore et al. reported that the pericardium mean dose and atria/ventricles volume covered by 45 Gy have detrimental effects on overall survival (12).

Clinical treatment planning is a trial-and-error process: the planner creates a treatment plan by iteratively updating optimization constraints in order to meet the target coverage while minimizing the dose to surrounding OARs. Despite best efforts from the planners, many clinical plans can be further improved in terms of OAR sparing. Such sub-optimal dose sparing can lead to increased doses to cardiac substructures and surrounding critical organs. In the past decade, knowledge-based planning (KBP) methods have been proposed to leverage knowledge embedded in prior high-quality plans to generate treatment plans for new patients (13, 14). Briefly, RapidPlan™ KBP estimates achievable DVHs for various structures included into the model, which are then translated into optimization objectives. The algorithm is divided into two parts, the model configuration and DVH estimation components. In model configuration, data extraction phase prepares the dataset for model training by extracting different geometric features of each structure. The model training phase produces a DVH estimation model for each structure. This is followed by generating estimated DVHs and optimization objectives through DVH estimation components. Previously, Harms et al. utilized RapidPlan™ KBP to improve cardiac sparing by incorporating the whole heart structure into the model training phase (15). While including the heart structure in the model would result in useful DVH predictions for the whole heart, it does not provide spatial information regarding dose distribution within the heart. We hypothesized that including cardiac substructures in the training phase of KBP would provide further dose reduction to critical cardiac substructures due to the additional spatial and geometrical features encoded into to each cardiac structure models. Therefore, the focus of this study was to build an improved KBP containing the features of each substructure (CS-KBP), in addition to other relevant OARs, to further enhance cardiac substructure sparing without compromising the target coverage. To assess the performance of a trained CS-KBP model, we used a validation dataset consisting of 28 additional patients that were not used in training the CS-KBP model.

We performed qualitative ( Figure 1 ) and quantitative ( Figures 2 , 3 , 4 ; S1 , 2 ) comparisons among the treatment plans generated by three methods. Figure 1 shows the dose distribution comparison among three different methods for two sample cases, each with the same CT scan slice, from the validation cohort. Clinically optimized plans are more conformal than the plans produced by C-KBP and CS-KBP for both cases, but at the expense of higher doses to coronary arteries such as LADA, LCFLX, and LMCA. While the C-KBP plan leads to better substructure sparing than clinical plan, the left circumflex coronary artery receives up to 33 Gy dose. CS-KBP plans spare cardiac substructures to a large extent with a tolerable tradeoff in plan conformity. Table 2 shows the dosimetric comparisons among the three techniques for the two cases shown in Figure 1 .

Figure 2A shows that the CS-KBP model significantly reduced the volume covered by 15, 20, 30, 40, and 45 Gy isodose volumes for LADA compared to two competing methods ( Figure 2B ). The CS-KBP model also produced plans with significantly lower mean and maximum doses (9.02 Gy and 20.04 Gy) than the plans generated by the C-KBP model (12.88 Gy and 25.94 Gy) and manual optimization (16.05 Gy and 29.65 Gy). A similar trend is observed for mean doses to LCFLX (12.74 Gy vs. 17.68 Gy and 20.68 Gy), LMCA (11.36 Gy vs. 20.44 Gy and 26.74 Gy), and RCA (11.36 Gy vs. 20.44 Gy and 26.74 Gy) ( Figures 2C, D ). In general, CS-KBP outperformed competing methods for reducing doses to coronary arteries.

Figure 3 outlines the doses to the chambers of the heart from the three techniques. While there was no significant differences in maximum dose and V_{63 Gy} for LA among the three techniques, CS-KBP produced plans with the significant sparing of LA measured in terms of mean dose (13.34 Gy vs. 16.22 Gy and 17.79 Gy) and V_{45 Gy} (6.74 cc vs. 7.88 cc and 11.13 cc) in comparison to treatment plans generated by the C-KBP model and manual optimizations ( Figures 3A, B ). For LV (3.55 Gy vs. 4.72 Gy and 6.52 Gy), RV (2.80 Gy vs. 4.11 Gy and 5.95 Gy) and RA (3.95 Gy vs. 5.59 Gy and 6.55 Gy), CS-KBP produced significant sparing across all dose volume metrics ( Figures 3C-F ).

Figure 4 summarizes the results for the great vessels across all three methods. Figure 4 shows a significant difference between CS-KBP and clinical plans for AA. However, Figure 4 shows that all three methods performed comparably (p > 0.05) for maximum doses to PA and SVC. However, CS-KBP plans resulted in significantly lower mean doses to all three great vessels compared to two competing methods. As shown in Figure S2 , CS-KBP also produced plans with significant sparing of all four cardiac valves compared to C-KBP and manually optimized plans. For whole heart structures, mean doses were 14.87 ± 5.01 Gy, 10.97 ± 3.41 Gy, and 8.95 ± 3.89 Gy for clinical, C-KBP, and CS-KBP plans, respectively. In terms of doses to the lungs, spinal cord, and esophagus, manually optimized plans largely performed better than both KBP models ( Figures S1 , S3 ). However, it is important to note that there was no violation of the clinically important dose volume constraints for these structures by treatment plans generated by the C-KBP and CS-KBP model.

A large number of studies in the literature have shown a correlation between post-radiotherapy cardiac toxicities and cardiac substructure irradiation (1, 3, 6, 10, 11, 29, 30). As a result, recent treatment planning efforts for thoracic cancers have been placed on sparing cardiac substructures surrounding the PTV compared to the traditional approach of sparing the heart as a single organ. Given a large number of substructures, it can be time-consuming to reach a desired treatment plan while effectively sparing substructures and other OARs. In general, it is challenging to build a KBP model for lung cancer due to variations in the location, shape, size, and orientation of PTV with respect to cardiac substructures and other critical OARs. This study developed a CS-KBP model and demonstrated its utility in the significant sparing of cardiac substructures for the VMAT of patients treated with definitive RT for locally advanced NSCLC.

For thoracic cancers, original QUANTEC recommendations are to limit the volume of the heart receiving ≥25 Gy less than 10% to keep the risk of cardiac mortality under 1% (31). Most current protocols now, however, aim to keep the mean heart dose below 20 Gy due to its correlation with a significantly higher rate of cardiac events (32). Recent works recommended substructure-specific dose volume constraints for various endpoints. For instance, D _max of 19.5 Gy to RCA has been associated with worse overall survival (11). The CS-KBP model (D _max = 8.1 ± 10.2 Gy) was able to reduce doses to RCA by 5.1 Gy and 10.1 Gy, respectively, compared to plans generated by C-KBP and manual optimization. For LADA, D _mean ≥23.8 Gy and V _{15 Gy} ≥10.0% has been associated with an increased percentage of serum troponin-T, which is an indicator of myocardial infarction. The CS-KBP model was able to generate plans with significantly lower mean doses to LADA compared to two competing methods (9.02 Gy vs. 12.88 Gy and 16.05 Gy), and also a lower volume of LADA covered with 15 Gy isodose than two competing methods (0.69 cc vs. 1.04 cc and 1.23 cc). In terms of great vessels, D _mean ≥8.5 Gy to AA, PA, and SVC has been linked to worse overall survival (9). While all three methods met these constraints, the CS-KBP model significantly reduced radiation doses to great vessels ( Figure 4 ). However, it is important to note that great vessels of the heart were not entirely contoured within the whole heart structure for clinical and C-KBP plans, which is one limitation of this study. For LV, the volume covered by the 5 Gy (V5) isodose has been linked to overall survival and acute coronary events (12, 33). The CS-KBP model-produced plans significantly improved this constraint compared to the C-KBP and optimized plans ( Figure 3 ). For LV, D _mean ≥ 6.7 Gy has been shown to increase the percentage of serum troponin (8). While all three techniques satisfied this constraint, the CS-KBP model resulted in lower D _mean to LV than C-KBP and manual optimization (3.55 Gy vs. 6.52 Gy and 4.72 Gy). For LV, V30 _Gy is also linked with symptomatic cardiac events (34). Plans generated by the CS-KBP model resulted in lower values of V30 _Gy compared to two competing methods (1.54 cc vs. 3.46 cc and 9.44 cc). In terms of atrial valve (AV), D _max ≥19.5 Gy has been associated with worse overall survival (11). While all three methods met this D _max constraint for AV, the CS-KBP model produced plans with the significantly lower D _max and D _mean to the valves of the heart ( Figure S2 ). Pericardial effusion is commonly reported as one of the cardiac toxicities post thoracic radiotherapy (35). The occurrence of pericardial effusion of any grade has been reported to be higher in patients with a mean heart dose >23.45 Gy (2). While the mean heart dose constraint is met with all three techniques in this study, the proposed CS-KBP outperforms the other two techniques (Clinical: 10.68 Gy; C-KBP: 8.57 Gy; CS-KBP: 6.36 Gy).

Sparing of cardiac substructures comes with some necessary tradeoffs. Radiation doses to 65% of the lungs were significantly lower in manually optimized plans than both KBP plans. In terms of mean doses to lungs, plans generated by C-KBP (14.23 Gy) resulted in slightly lower doses than CS-KBP (14.85 Gy) and manual optimization (14.31 Gy). In terms of the dose received by 0.03 cc of the spinal cord, manually optimized plans resulted in the lowest dose (28.06 Gy) compared to C-KBP (30.77 Gy) and CS-KBP (34.26 Gy). Nonetheless, all three techniques satisfied the clinical constraints of keeping the maximum cord dose below 40 Gy. Plans generated with CS-KBP had higher plan complexity (630MU) than the ones generated via the C-KBP model (565MU) and manual optimization (486 MU). This tradeoff among different structures and in increased plan complexity is consistent with that of Harms et al., who found that their clinical RapidPlan™ (RP) model significantly outperformed the cardiac sparing RP model with increased plan complexity by the cardiac sparing RP model in terms of the number of monitor units (15). While CS-KBP led to the lowest overall doses to the cardiac substructures, the two competing methods also achieved the dose-volume constraints. The clinical impact of recently introduced cardiac substructure-specific dose constraints is actively being studied based on randomized trials and from real-world datasets on lung cancer (36). In contrast, there are standardized dose volume constraints for other OARs (i.e., lung and spinal cord). This study demonstrates the ability of CS-KBP to meet these standard dose volume constraints of other OARs while significantly reducing the doses to cardiac substructures. Though the direct clinical benefits from reducing the dose to cardiac substructures is yet to be validated with patient outcome data, lower doses to these cardiac substructures are expected to translate to reduced probability of radiation-induced cardiac toxicities post radiotherapy.

A limitation of the present study is that all the manually optimized plans were generated by a few dosimetrists and all thoracic OARs were contoured by a single observer. Another limitation of this study is the use of conventional CT for contouring cardiac substructures, which may introduce uncertainty due to the poor discernibility of substructures on CT images. This limitation can potentially be addressed by previous studies that leveraged the enhanced soft tissue contrast of MRI with conventional non-contrast CT to develop a deep learning-based framework (37) and a hybrid MR/CT segmentation atlas-based framework (38) for cardiac substructure segmentation. Though training the sample size was relatively small, it was within a range that had been shown to be adequate for a KBP model for thoracic radiation therapy (39). Another minor limitation of this study is a slight mismatch between the contours and the plans. Cardiac substructures were contoured on free-breathing scans for higher spatial resolution and then transferred via rigid registration to the average 4D CT for treatment planning. Although all patients in the cohort were treated with a free breathing technique, slight differences between the free breathing scans and the 4D average scans can be expected. Nonetheless, since all three sets of plans were generated on the same datasets, the results reflect the true differences among the three techniques. Following preliminary cardiac substructure segmentations by a physician resident, all cardiac substructures were verified by an attending physician. While we did not perform a robust evaluation for the quality of segmentation due to the nature of this study; different strategies such as percent overlap of observer volumes, a robust atlas-based method, etc. can be utilized clinically to ensure high quality segmentation of cardiac substructures.

All treatment plans were generated within one institution, lacking variations in terms of contouring and dosimetric protocols from multi-institutions. A validation cohort from multi-institutional data is necessary to further ensure the robustness of this model and also consider differences between planners and dosimetry protocols. Nonetheless, there was a large variation in the size of the PTV volumes in the validation cohort (454.05 ± 249.42 cc) of this study, demonstrating the feasibility of the proposed model for variable tumor sizes. In this study, all the structures were manually contoured, which may be time consuming. Various deep learning frameworks have been introduced to increase treatment planning efficiency (40). The impact of auto-contouring versus manual contouring on the prediction model will be investigated in the future studies. While KBP models have been built for various treatment sites, it has not been widely adopted for the purpose of cardiac substructure sparing (13). Overall results demonstrated the ability of proposed CS-KBP model to achieve significant cardiac substructure sparing compared to the traditional method of manual optimization as well as the KBP model based on cardiac structure alone (C-KBP).

The proposed approach can aid future researchers to implement the knowledge-based planning module to effectively spare cardiac substructures while satisfying the clinical constraints of other OARs and target coverage and also further enhance treatment planning efficiency. This model may offer reduced planning time, improve plan quality, and should be studied in the prospective setting as a way to improve outcomes in patients with locally advanced lung cancer.

The original contributions presented in the study are included in the article/ Supplementary Material . Further inquiries can be directed to the corresponding author.

SM: First authorship. JW, JR, YL, TL, JB, KH, XY: Co-authorship. JZ: Senior authorship. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.