Optimizing allocation of limited proton fractions via NTCP-optimized combined proton–photon radiotherapy

Abstract

Compared to photon radiotherapy (RT), proton RT is less widely available and more costly. To maximize the normal-tissue complication probability (NTCP) benefit of limited proton resources at a population level, this work introduces a novel NTCP-optimized combined proton–photon treatment (NTCP-CPPT) approach that can achieve the OAR-sparing benefits in terms of reaching desired NTCP thresholds with the minimal proton fraction. NTCP-optimized intensity-modulated proton therapy (NTCP-IMPT) and NTCP-optimized intensity-modulated radiation therapy (NTCP-IMRT) plans are generated via the NTCP-optimized treatment planning based on physical dose and NTCP objectives. For each patient, a minimal proton fraction is determined for the NTCP-CPPT plan as a combination of NTCP-IMPT and NTCP-IMRT plans, while still retaining the sum-NTCP threshold for the patient to be justified for the proton access. For comparison, conventional CPPT (CONV-CPPT) plans are obtained by a combination of conventional IMRT (CONV-IMRT) and conventional IMPT (CONV-IMPT) plans that are solely optimized based on physical dose optimization. Compared to CONV-CPPT, NTCP-CPPT reduced the average proton fractions from 18 to 3 without compromising physical dose objectives, while both NTCP-CPPT and CONV-CPPT plans maintained a sum-NTCP below 20% relative to CONV-IMRT. Additionally, NTCP-CPPT, using an average of 22 proton fractions out of 35, maintained the same sum-NTCP as CONV-IMPT required 35 proton fractions. To increase patient access to proton RT, a novel combined proton–photon treatment called NTCP-CPPT is proposed that can provide a minimal number of proton fractions while still retaining sufficient OAR-sparing benefits to justify the proton access.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s13246-026-01759-7.

Introduction

Cancer remains a leading cause of mortality worldwide, with over 50% of cancer patients receiving radiation therapy (RT) [1]. Proton RT is regarded as an advanced treatment because it delivers a lower integral dose at the proximal and minimizes the distal dose to organs at risk (OAR) compared to conventional photon RT [2–4]. However, access to proton RT is hindered by limited proton resources and high treatment costs. Fewer than 1% of cancer patients globally receive proton RT, although moderate estimates suggest that 15 to 50% of patients could potentially benefit from it [5–7].

Many countries have now established standards to determine patient eligibility for proton RT [8–10]. Patient selection based on the normal-tissue complication probability (NTCP) model has been adopted by Dutch health authorities for choosing between proton and photon RT [11–15]. Proton RT is recommended only when the NTCP reduction (NTCP_photon–NTCP_proton) exceeds a specified threshold. Due to the convex shape of NTCP curves, the NTCP marginal benefit does not consistently increase with an increasing number of proton fractions. A strategy known as combined proton–photon treatment (CPPT), where a portion of the fractions is delivered with protons and the remainder with photons, is employed to maximize the NTCP benefit on a population level and increase access to proton RT under limited resources.

The rationale of CPPT has been investigated in early clinical research [16–18]. CPPT has been employed in clinical practice to treat spine sarcoma, where photon and proton plans are optimized individually to ensure that each modality delivers the prescribed dose per fraction to the target volume [19]. Unkelbach et al. [20, 21] and Ten Eikelder et al. [22] developed a treatment planning method that simultaneously optimizes intensity-modulated radiation therapy (IMRT) and intensity-modulated proton therapy (IMPT) plans with a fixed fractionation scheme, using the biologically effective dose (BED) model to account for fractionation effects. Gao [23] conducted simultaneous optimization of proton weight and photon fluence based on cumulative physical dose. This method ensured that both proton and photon components achieved their respective dose objectives, allowing them to be delivered independently. Fabiano et al. [24] and Amstutz et al. [25] investigated methods for optimizing CPPT to handle the range and anatomic uncertainty, respectively. Fabiano et al. [26] proposed a novel approach to CPPT without a proton gantry, using a treatment room equipped with a fixed proton beamline, a standard linear accelerator (LINAC) for photon RT, and a robotic treatment couch, enabling the delivery of protons and photons in a single fraction.

Many studies have focused on optimizing dose quality and treatment deliverability, but the issue of how best to allocate limited proton fractions in combined treatments remains unresolved. Li et al. [27] developed a hybrid treatment planning method that concurrently optimizes the proton fraction, IMRT plan, and IMPT plan for better plan dose quality. Levin et al. [28] introduced a mathematical framework to simultaneously optimize dose distribution and fractionations of proton and photon modalities to maximize tumor control and spare the surrounding healthy tissue. However, these studies did not consider the restricted availability of proton resources. Loizeau et al. [29] optimally distributed a limited number of proton therapy slots for CPPT over a head and neck (HN) patient cohort to maximize NTCP benefits on a population level. In this study, the standard CPPT did not further optimize NTCP, and therefore, it offered no significant advantage in reducing the sum-NTCP across patient cohorts compared to patient selection strategies. There remains a need for a more efficient CPPT approach with fewer proton fractions while preserving comparable normal tissue toxicity.

This study aims to develop a novel NTCP-optimized CPPT (NTCP-CPPT) approach that optimizes proton fractions for each patient to achieve the OAR-sparing benefits in terms of reaching the desired sum-NTCP thresholds with a minimal number of proton fractions. NTCP-optimized IMPT (NTCP-IMPT) and NTCP-optimized IMRT (NTCP-IMRT) plans are created through NTCP-optimized treatment planning, accounting for both physical dose and NTCP objectives. The NTCP-CPPT plan is the combination of NTCP-IMPT and NTCP-IMRT plans, whose proton fractions are determined based on the desired sum-NTCP threshold. Finally, the NTCP-CPPT plan is compared with the conventional CPPT (CONV-CPPT) regarding physical dose objective, NTCP objective, and proton fractions.

Methods

Patients and dose prescription

To assess the effectiveness of the proposed method, retrospective treatment planning was performed for 12 head and neck (HN) patients previously treated with IMPT. This research was carried out under Human Subject Assurance Number 00003411 for University of Kansas in accordance with the principles embodied in the Declaration of Helsinki and in accordance with local statutory requirements. The treatment plan utilized a simultaneous integrated boost (SIB) technique, prescribing two different dose-level: 59.40 GyRBE to the clinical target volume (CTV5940) and 69.96 GyRBE to the boost clinical target volume (CTV6996) over 35 fractions. The dose constraints (Table 1) were established in part according to the experience of Widesott et al. [30] and in part according to our own experience [31]. A constant relative biological effectiveness (RBE) of 1.1 was applied in the proton treatment plan. Detailed patient characteristics and plan settings are provided in Table 1.

Table 1.

Patient characteristics and plan setting

Patient number	Target volume (cc)			Target position	Proton beam angle	Number of photon beam
	CTV6996	CTV5940	PTV5940
1	76.6	417.3	1162.6	Bilateral	[0°, 30°, − 30°, 120°, − 120°]	16
2	54.2	338.2	779.9	Right	[− 60°, 0°, − 120°]	16
3	65.5	497.0	1135.1	Bilateral	[0°, 45°, − 45°, 140°, − 140°]	16
4	206.8	536.1	1250.2	Bilateral	[0°, 45°, − 45°, 140°, − 140°]	16
5	104.3	380.3	1069.2	Bilateral	[0°, 45°, − 45°, 140°, − 140°]	16
6	78.0	281.6	806.7	Left	[0°, 45°, 90°]	16
7	40.1	183.2	623.2	Bilateral	[0°, 45°, − 45°, 140°, − 140°]	16
8	70.1	184.5	469.5	Right	[5°, − 30°, − 90°]	16
9	242.1	490.1	908.1	Left	[− 45°, 45°, 135°]	16
10	103.8	234.1	531.7	Right	[5°, − 30°, − 90°]	16
11	34.5	323.8	802.3	Bilateral	[0°, 45°, − 45°, 140°, − 140°]	16
12	238.8	345.6	697.3	Right	[− 60°, 0°, − 120°]	16

Conventional treatment planning

The CONV-IMPT planning considered the plan dose robustness and deliverable minimum-monitor-unit (MMU) constraint. The dose fidelity term was used to achieve the clinical dose goal by minimizing the dose objective. Apart from the dose objectives, the MMU regularization term was added to the optimization objective function to ensure plan deliverability, as shown in Eq. (1–5).

1

2

3

4

x ≥ 0 5

where f(d) is the dose fidelity term, g(x) is the MMU regularization term, x is the variable vector and α is the weight of the MMU regularization term. The scenario weights p_k are defined as the probability that scenario k occurs and A_k is the responding dose influence matrix. And x_min is the MMU. As demonstrated in Table S1, there are five types of dose fidelity objectives: squared dose deviation, squared overdosing deviation, squared underdosing deviation, DVH-max dose deviation, and DVH-min dose deviation.

For CONV-IMPT treatment planning, the stochastic robust optimization was used to ensure the robust dose distribution accounting for range and setup uncertainties. This approach includes 9 scenarios, with a 3 mm isotropic setup uncertainty and a 3.5% range uncertainty for all cases [32]. Also, the number of coplanar angles ranged from 3 to 5, as indicated in Table 1. The beam angle arrangement was adapted to the specific anatomy of each patient and the 4 cm range shift was used to ensure that the tumor was adequately covered, if necessary. The candidate spots were selected over the target using a rectangular grid with a lateral spacing of 5 mm and longitudinal spacing of 5 mm. For the CONV-IMRT treatment planning, PTV-based optimization was used. Specifically, PTV5940 was generated by applying a 5 mm margin to CTV5940 to account for setup uncertainties. No PTV was defined for CTV6996, and dose coverage for CTV6996 was directly optimized. We consider an IMRT plan consisting of 16 coplanar beams, which approximates the best possible plan that can be delivered with volumetric modulated arc therapy (VMAT). The dose distributions were calculated via matRad [33] on a 3 *3 *3 mm³ dose grid. Doses to OAR were minimized in line with the dose objectives in Table S2.

NTCP-optimized treatment planning

NTCP-optimized treatment planning simultaneously optimized physical dose quality, MMU, and normal-tissue toxicity. As shown in Eq. (6), in addition to the physical dose objectives and the MMU constraint, the weighted sum-NTCP was incorporated into the objective function to minimize normal-tissue toxicity.

6

where β is the regularization weight of the sum-NTCP term. The Netherlands LIPPv2.2 for xerostomia and dysphagia (Grade ≥ 2 and Grade ≥ 3) [34] for proton–photon patient selection was considered in the NTCP-optimized treatment planning. The NTCP values of different side effects were calculated as shown in Eq. (7).

7

where D_mean is the mean dose to OAR, with corresponding variables and regression coefficients detailed in Table S3. We used the mean dose of pharyngeal constrictor muscles (PCMs) as a substitute for the superior, medius, and inferior parts, which were not separately delineated at our institute.

Combined proton–photon planning

Initially, single-modality 35-fraction CONV-IMRT and CONV-IMPT plans were optimized using conventional treatment planning. Single-modality 35-fraction NTCP-IMRT and NTCP-IMPT plans were then created using NTCP-optimized treatment planning. The NTCP-CPPT plan was developed by proportionally combining NTCP-IMPT and NTCP-IMRT plans (i.e., delivering NTCP-IMPT n_p fractions and NTCP-IMRT n_r fractions with n_p + n_r = 35). Similarly, the CONV-CPPT plan was created by combining the CONV-IMPT and CONV-IMRT plans.

This goal is to minimize proton fractions for each patient to achieve the OAR-sparing benefits in terms of reaching desired sum-NTCP thresholds with a minimal number of proton fractions. In our study, the optimal proton fraction of NTCP-CPPT and CONV-CPPT was determined by the threshold-determined method. The problem is formulated as follows:

8

here, n_p is the proton fraction in the NTCP-CPPT or CONV-CPPT plan. NTCP_CPPT and NTCP_CONV-IMRT refer to the sum-NTCP of different side effects for the CPPT and CONV-IMRT plans, respectively. We used a metric, termed sum-NTCP, which represents the equally weighted sum of the NTCP values for xerostomia and dysphagia as shown in Table S3. It is important to note that sum-NTCP does not directly represent a probability but serves as an evaluation metric and is also used for NTCP-based patient selection. In this study, the threshold is 20%. This indicates that the NTCP reduction (NTCP_CONV-IMRT – NTCP_CPPT) for each patient must exceed the 20% threshold.

We also selected the minimal proton fraction of the NTCP-CPPT to achieve normal-tissue toxicity comparable to that of CONV-IMPT.

9

where the NTCP_CONV-IMPT refers to the sum-NTCP of different side effects of the CONV-IMPT plan.

Evaluation of treatment plans

For each patient, single-modality NTCP-IMPT and NTCP-IMRT plans were generated by adjusting the regularization weight (β) of the NTCP term in the NTCP-optimized treatment planning to balance physical dose quality and sum-NTCP. The sum-NTCP and dosimetric parameters of these plans were compared to those of CONV-IMPT and CONV-IMRT.

The NTCP-CPPT plan was created by combining NTCP-IMRT and NTCP-IMPT with minimal proton fraction to achieve OAR-sparing benefits. The number of proton fractions required to achieve similar NTCP thresholds for NTCP-CPPT and CONV-CPPT were compared. Physical dose parameters for CTV6996 (D_2% and D_98%), CTV5940 (D_2% and D_98%), spinal cord (D_2%), and brain stem (D_2%) were also evaluated.

Results

NTCP-optimized treatment planning

For all cases, CONV-IMRT and CONV-IMPT plans were generated using conventional treatment planning based on physical dose objectives, while NTCP-IMRT and NTCP-IMPT plans were created based on the physical dose and NTCP objectives. By adjusting the regularization weight (β) of the NTCP term, a set of plans was created for each patient. As shown in Fig. 1, NTCP-IMRT and NTCP-IMPT plans were selected where the dose objective value f was below 1.1 times f_CONV-IMPT and f_CONV-IMRT, respectively.

Fig. 1.

The generation of NTCP-IMRT (a) and NTCP-IMPT plans (b) using NTCP-optimized treatment planning based on the physical dose and NTCP objectives. Each data point represents one plan

NTCP benefit for NTCP-optimized treatment planning

Due to the OAR-sparing benefits of protons, CONV-IMPT reduced the sum-NTCP from 101.8% (range: 63.9%–132.3%) to 60.7% (range: 39.2–96.1%) on average compared to CONV-IMRT, as shown in Fig. 2. When the NTCP objective was included into the optimization objective, NTCP-IMPT also lowered the sum-NTCP from 80.8% (range: 53.9–115.2%) to 50.3% (range: 31.6–77.9%) in comparison to NTCP-IMRT. This demonstrated that including normal tissue toxicity (sum-NTCP) optimization alters dose distribution and may reduce toxicity. NTCP-IMRT reduced sum-NTCP by 20.4% (range: 10.1–28.0%) over CONV-IMRT, and NTCP-IMPT lowered it by 10.4% (range: 4.1–21.4%) over CONV-IMPT.

Fig. 2.

Normal tissue complication probability (Dysphagia Grade ≥ 2, Dysphagia Grade ≥ 3, Xerostomia Grade ≥ 2, and Xerostomia Grade ≥ 3) for the 12 HN cancer patients for the CONV-IMRT, CONV-IMPT, NTCP-IMRT, and NTCP-IMPT plans

Figure 2 indicates that a reduction in the sum-NTCP was feasible, but it varied among patients. Figure 3 demonstrates that NTCP-optimized treatment planning achieved mean dose reductions in the parotid, submandibular, oral cavity, and PCMs compared to conventional treatment planning. Other dosimetric parameters, including CTV D_2%, D_98%, HI, and maximum doses to the spinal cord and brain stem, showed no significant differences between conventional and NTCP-optimized treatment planning.

Fig. 3.

The dose map of CONV-IMRT (a), NTCP-IMRT (b), CONV-IMPT (d), NTCP-IMPT (e), and the dose difference map between CONV-IMRT and NTCP-IMRT (c), CONV-IMPT and NTCP-IMPT (f)

Proton fraction benefit of NTCP-CPPT over CONV-CPPT

As shown in Fig. 4, NTCP-CPPT significantly reduced the required number of proton fractions from an average of 18 (range: 10–28) to 3 (range: 0–15) across all patients than CONV-CPPT. Both NTCP-CPPT and CONV-CPPT plans maintained a sum-NTCP below 20% relative to the CONV-IMRT plan. As illustrated in Fig. 5(a–c), the target dose quality was comparable between CONV-CPPT with 21 proton fractions and NTCP-CPPT with 4 proton fractions for the HN patients, while the CONV-CPPT and NTCP-CPPT have the same sum-NTCP. Similarly, Fig. 6(a, b) shows no significant difference in dose-volume metrics, with NTCP-CPPT plans yielding similar hot spots (CTV6996 D_2%: 73.1 (range: 72.0–74.6) vs. 73.2 (range: 72.1–74.6) GyRBE) and cold spots (CTV6996 D_98%: 69.5 (range: 69.1–69.7) vs. 69.4 (range: 69.2–69.8) GyRBE). However, with increased photon fractions, the D_2% for the brain stem (27.3 (range: 2.8–39.6) vs. 19.1 (range: 1.5–34.6) GyRBE) and spinal cord (35.7 (range: 30.3–41.8) vs. 25.6 (range: 20.0–32.6) GyRBE) was higher in NTCP-CPPT (Fig. 6(c, d)), though still satisfied clinical goals (Table S2).

Fig. 4.

The comparison of NTCP (Dysphagia Grade ≥ 2, Dysphagia Grade ≥ 3, Xerostomia Grade ≥ 2, and Xerostomia Grade ≥ 3) for the 12 HN cancer patients between the CONV-CPPT and NTCP-CPPT plans (a). The comparison of needed proton fractions for the 12 HN cancer patients between the CONV-CPPT and NTCP-CPPT plans (b)

Fig. 5.

The dose map of CONV-CPPT with 21 CONV-IMPT fractions (a), NTCP-CPPT with 4 NTCP-IMPT fractions (b), and the dose difference map between the CONV-CPPT and the NTCP-CPPT (c) for the selected HN patient. The dose map of CONV-IMPT with 35 proton fractions (d), NTCP-CPPT with 16 NTCP-IMPT fractions (e), and the dose difference map between the CONV-IMPT and the NTCP-CPPT (f) for the selected HN patient

Fig. 6.

The dosimetric comparison of CTV6996 D_2% (a), CTV699 D_98% (b), spinal cord D_2% (c), and brain stem D_2% (d) for the CONV-IMRT, CONV-CPPT, and NTCP-CPPT plans for the selected HN patient

Proton fraction benefit of NTCP-CPPT over CONV-IMPT

We also explored the potential to reduce proton fractions in NTCP-CPPT while maintaining normal-tissue toxicity comparable to CONV-IMPT. As shown in Fig. 7, NTCP-CPPT required only 22 proton fractions (range: 16–31) to achieve a similar sum-NTCP as CONV-IMPT, which typically uses 35. In Fig. 5(d–f), target dose quality for HN patients remained comparable between CONV-IMPT (35 proton fractions) and NTCP-CPPT (16 proton fractions). However, NTCP-CPPT showed higher D_2% for the brain stem and spinal cord compared to CONV-IMPT, as indicated in Fig. 8 (c, d).

Fig. 7.

The comparison of NTCP (Dysphagia Grade ≥ 2, Dysphagia Grade ≥ 3, Xerostomia Grade ≥ 2, and Xerostomia Grade ≥ 3) for the 12 HN cancer patients between the CONV-IMPT and NTCP-CPPT plans (a). The comparison of needed proton fractions for the 12 HN cancer patients between the CONV-IMPT and NTCP-CPPT plans (b)

Fig. 8.

The dosimetric comparison of CTV6996 D_2% (a), CTV6996 D_98% (b), spinal cord D_2% (c), and brain stem D_2% (d) for the CONV-IMPT, NTCP-CPPT plans

Discussion

Proton therapy centers are costly to establish and maintain, leading to limited proton RT slots compared to conventional photon RT. To address this limitation, combined proton–photon treatments have been investigated to use these scarce proton slots. A key question is how to achieve comparable treatment outcomes with fewer proton fractions. This study developed an NTCP-optimized combined proton–photon treatment approach, selecting the optimal proton fraction to maximize OAR-sparing benefits in terms of achieving the desired sum-NTCP while efficiently utilizing limited proton resources.

We proposed implementing the NTCP-based treatment planning approach to reduce post-treatment toxicity. This method incorporated NTCP model-based objectives for OARs related to toxicity, apart from the physical dose objectives used in conventional treatment planning. The approach generated NTCP-IMPT and NTCP-IMRT plans with comparable target coverage and improved OAR sparing versus CONV-IMRT and CONV-IMPT plans. Significant NTCP reductions for xerostomia and dysphagia were observed due to lower mean doses of related OARs. Furthermore, NTCP-CPPT effectively reduced the required number of proton fractions from an average of 18 to 3 across all patients than CONV-CPPT while both plans maintained a sum-NTCP that was greater than 20% less than the sum-NTCP for CONV-IMRT. The 20% sum-NTCP threshold is regarded as the threshold in the NTCP-based patient selection between proton and photon RT. Although, patients qualify for proton therapy in the Netherlands in case the ΔNTCP (NTCP_photon—NTCP_proton) ≥ 10% for grade ≥ 2, ≥ 5% for grade ≥ 3 side effects, or the total NTCP reduction exceeds 15% [12]. If this 15% threshold is applied to our study, a further reduction in proton fraction will be observed. It is particularly significant that, as shown in Fig. 4b, when NTCP optimization was applied to the CPPT plan, 5 out of 12 HN patients did not require proton fractions. This highlights the significant potential of incorporating NTCP optimization into photon treatment planning, demonstrating its value in reducing normal tissue toxicities. Although NTCP-based optimization holds clear promise, its clinical use remains limited. A major barrier is the lack of standardized and broadly validated NTCP models. Most existing models are developed from specific patient cohorts or treatment techniques, and their parameters may not generalize reliably across institutions or patient populations, leading to uncertainty in the predicted toxicity risks.

The cost of proton RT is typically three times higher than photon RT [35, 36]. Limited insurance coverage and high treatment costs often restrict access to proton therapy, particularly for low- and middle-income populations. Using fewer proton fractions in NTCP-CPPT can reduce the treatment expense while maintaining the dose quality. Several studies have sought to lower treatment expenses by improving treatment efficiency and efficacy [37–39] and decreasing the size of the proton therapy machine with super-conducting gantry [40] and gantry-less proton therapy schemes [41]. A plausible hypothesis is that a more accessible and affordable proton therapy option could result in a modest yet significant decrease in sum-NTCP values for a specific group of patients, considering the limited proton resources.

This study has several limitations. First, in this study, robust optimization (3 mm/3.5%) was applied to address uncertainty for IMPT, while for IMRT, a 5 mm PTV margin was used to account for setup uncertainties. The differing approaches align with clinical practices. However, it is important to note that the size of the PTV margin used in IMRT treatment planning can significantly affect the required proton fractions to achieve the NTCP threshold in both CONV-CPPT and NTCP-CPPT plans. Also, the proposed method separated IMPT and IMRT planning to allow for delivery in different fractions. Due to the varying proton and photon fractions, future studies should consider the fractionation effect to optimize the biologically effective dose. Integrating fractionation, biologically effective dose, and NTCP objectives optimization into a unified framework is a goal for future work.

In this study, to generate the photon reference plans, a 16-field coplanar IMRT plan was used. However, in routine clinical workflows, more advanced delivery techniques VMAT are frequently preferred. The decision to use a 16-beam IMRT plan in this study was driven by constraints of the matRad planning platform, which currently lacks comprehensive functionality for VMAT planning, including the optimization of continuously rotating gantry angles and dynamically varying MLC leaf trajectories. It should be noted that a clinically generated VMAT plan might offer superior dosimetric quality relative to the IMRT plans employed here.

Step-and-shoot IMRT plans with 16 fields and IMPT plans with 3–5 beams were used in this study. Increasing the number of beams is expected to improve both physical dose quality and NTCP benefits. The proposed NTCP-based treatment planning approach optimizes proton spot weight and photon fluence to minimize NTCP. However, NTCP also depends on beam angle selection, which was predefined in this work. Further investigation is needed to integrate beam angle optimization [42, 43] into the NTCP optimization framework to further reduce radiation-induced toxicity.

Conclusion

In conclusion, for the head and neck patient cohort evaluated in this study, the proposed NTCP-CPPT planning framework effectively reduces the required number of proton fractions to achieve comparable sum-NTCP values than the CONV-CPPT planning method. By incorporating NTCP-based optimization framework into treatment planning, this approach has the potential to reduce overall treatment costs, thereby improving patient access to proton radiotherapy.

Supplementary Information

Below is the link to the electronic supplementary material.

Funding

Foundation for the National Institutes of Health, R37CA250921, Hao Gao, R01CA261964, Hao Gao.

Data Availability

Up on reasonable request data utilized in this study will be provided.

Declarations

Conflict of interest

The authors declare no conflict of interest.

Ethics approval

This retrospective study was approved by the Institutional Review Board of UT Southwestern Medical Center. The requirement for informed consent was waived because the study involved retrospective analysis of existing de-identified clinical imaging data.