Skip to main content
Download PDF

A deep learning-informed interpretation of why and when dose metrics outside the PTV can affect the risk of distant metastasis in SBRT NSCLC patients

Radiation Oncology volume 19, Article number: 127 (2024) Cite this article

Abstract

Purpose

Recent papers suggested a correlation between the risk of distant metastasis (DM) and dose outside the PTV, though conclusions in different publications conflicted. This study resolves these conflicts and provides a compelling explanation of prognostic factors.

Materials and methods

A dataset of 478 NSCLC patients treated with SBRT (IMRT or VMAT) was analyzed. We developed a deep learning model for DM prediction and explainable AI was used to identify the most significant prognostic factors. Subsequently, the prognostic power of the extracted features and clinical details were analyzed using conventional statistical methods.

Results

Treatment technique, tumor features, and dosiomic features in a 3 cm wide ring around the PTV (PTV3cm) were identified as the strongest predictors of DM. The Hazard Ratio (HR) for Dmean,PTV3cm was significantly above 1 (p < 0.001). There was no significance of the PTV3cm dose after treatment technique stratification. However, the dose in PTV3cm was found to be a highly significant DM predictor (HR > 1, p = 0.004) when analyzing only VMAT patients with small and spherical tumors (i.e., sphericity > 0.5).

Conclusions

The main reason for conflicting conclusions in previous papers was inconsistent datasets and insufficient consideration of confounding variables. No causal correlation between the risk of DM and dose outside the PTV was found. However, the mean dose to PTV3cm can be a significant predictor of DM in small spherical targets treated with VMAT, which might clinically imply considering larger PTV margins for smaller, more spherical tumors (e.g., if IGTV > 2 cm, then margin ≤ 7 mm, else margin > 7 mm).

Introduction

Non-small cell lung cancer (NSCLC) is responsible for the highest number of cancer-related deaths in the European Union [1]. One of the standard treatment techniques for early-stage NSCLC is stereotactic body radiotherapy (SBRT). In recent years, several publications have described the impact of dose outside the planning target volume (PTV) on the probability of distant metastasis (DM) [2,3,4,5,6,7]. Unfortunately, the conclusions were conflicting, leading to confusion around this issue.

The first study, by Diamant et al., [2] demonstrated a higher incidence of distant metastasis with a lower mean dose to a 3 cm shell region around PTV. This challenged conventional practice that highly conformal dose distribution should be strictly limited to the PTV, with steep dose fall-off. A subsequent paper reported much better outcomes of CyberKnife (CK) in terms of distant metastasis compared to 3D conformal radiotherapy (3D-CRT) and volumetric arc therapy (VMAT), using a similar argument [3].

Another study, by Hughes et al., aimed to validate Diamant’s findings and found no significant association between dose outside PTV and DM [6]. Lalonde et al., on the other hand, found exactly the opposite correlation from Diamant, with patients receiving lower mean doses to the 3 cm ring experiencing reduced risk of DM.

In this study, we aimed to develop a discrete-time interval outcome model [8, 9], incorporating multi-modality features from CT data, 3D planning dose distribution, and patient clinical/demographic details to predict distant metastasis in NSCLC patients treated with SBRT. We identified regions in the planning dose distribution critical to the model’s decision-making using modern deep learning explainability methods. These results, combined with a thorough statistical analysis of our dataset, allowed us to reconcile previously reported conflicting conclusions from the literature concerning the impact of dose outside the PTV on the risk of DM [2,3,4, 6].

Material and methods

Patient dataset

We analyzed 478 sequential non-small cell lung cancer (NSCLC) patients treated with SBRT between 2010 and 2017. Only patients with PET/CT staging and no prior RT treatment at our institution were included. For consistency, we only included patients treated with a prescription dose of 50 Gy in 5 fractions (BED = 100 Gy). The prescription required 95% of the PTV and 100% of the IGTV to be covered by 50 Gy. In a limited number of cases, these requirements were slightly relaxed to spare normal tissues directly abutting the tumor. Patient characteristics are provided in Table 1.

Table 1 Patient characteristics
Full size table

Tumor segmentation and treatment planning

All patients underwent 4D CT scanning with 10 equally spaced respiratory phases. The average CT scan was utilized for treatment planning. These scans were obtained with a resolution of 512 × 512 and a slice thickness of 3 mm (1.17 × 1.17 × 3 mm3). Tumor delineation was performed using the IGTV technique [10]. Tumors were contoured in lung windows and included tumor spiculations within the IGTV. PTV structures were usually created as expansions from IGTV with a 5 mm margin radially and a 7 mm margin superiorly/inferiorly. A minority of patients had a concentric 5 mm PTV margin.

Treatment planning of all patients was provided in Pinnacle TPS (v8.0–v16.2) using a convolution/superposition dose calculation algorithm. The resolution of the calculated 3D dose distribution was mostly 3 × 3 × 3 mm3. Two SBRT treatment techniques- static IMRT and VMAT- were employed within the dataset. All treatments were delivered on Varian TrueBeam or Varian Trilogy linear accelerators. The average PTV Dmedian was 109.6 Gy BED (IQR: 106.7–112.3 Gy BED), the average PTV Dnear-min (D98) 94.3 Gy BED (IQR: 94.4–98.4 Gy BED), and the average PTV Dnear-max (D2) 119.0 Gy BED (IQR: 112.9–123.8 Gy BED).

Follow-up and clinical endpoints

Follow-up appointments were typically scheduled every 3 months in the first year after treatment and every 6 months for the next 4 years.

The primary endpoint for this study was distant metastasis, as in previously published papers [2,3,4, 6]. The rate of DM in our cohort was 19%. Rates of other clinical endpoints (5 years after the treatment) were—locoregional recurrence (LRR—23%), local recurrence (LR—6%), and overall survival (OS—58%). There was an overlap of 41 patients between DM and LRR patients. Median follow-up, including interquartile ranges (IQR), can be found at the bottom of Table 1.

Deep learning survival model

The deep-learning survival model (DL-surv) is designed as a featureless model that accepts multi-modal data at the input—planning CT slices, planning 3D dose distribution, and patient clinical/demographic details, such as gender, age, PTV volume, clinical maximum tumor diameter (CMD) or lobe of the primary tumor. Its diagram is shown in Fig. 1. It consists of three modules for dimensionality reduction and feature extraction in the first stage, and one survival neural network (Surv-nnet) in the second stage, which simulates DM-free conditional probabilities in discrete-time intervals (0–1, 1–2, 2–3 and 3–5 years). Such a design is inspired by models utilized in earlier studies [8, 9, 11, 12]. More details, including architecture diagrams and data preprocessing, are provided in Appendix A.

Fig. 1
figure 1

Diagram of DL-surv model for time to distant metastasis prediction

Full size image

Technical details of Surv-nnet and its loss function have been thoroughly described previously [8]. Details, including a description of the DL-surv loss function, are provided in Appendix B.

The dataset was divided into training and testing cohorts following the TRIPOD criteria 2b [13]. The cohorts were stratified by time-to-DM with 20% of the samples comprising the testing subset, which was left untouched for independent testing of the final model. The remaining 80% (training subset) was employed in hyperparameter (HP) tuning and fivefold cross-validation (CV) with 50 iterations in total. Among the tuned HP were learning rate (lr), batch size (bs), weight decay (wd), dropout ratio (do), wDemo-VAE, and minority class weight (wc). The final model was trained using optimized hyperparameters: lr = 1e−4; bs = 32; wd = 0.1; do = 0.2; wDemo-VAE = 1; wc = 1. The model’s performance was evaluated using Harrell’s concordance index (c-index) [14], which was evaluated at the last time interval.

The Gradient-weighted Class Activation Mapping (Grad-CAM) method [15] was used to provide a graphical interpretation of the DL-surv model’s decision-making and explanation of influencing features regarding distant metastasis probability. It was applied after the last convolutional layer in Dose-CNN (see Fig. 1).

Statistical evaluation and dose parameters

The main aim of this work was to determine and explain significant dose metrics, tumor features, or other parameters in the treatment plans regarding the risk of DM, and thereby explain the divergent conclusions from the previous studies. Therefore, appropriate metrics were studied considering dataset characteristics and Grad-CAM results. Typical dose metrics, such as Dmean, Dnear-min and Dnear-max, were analyzed in the IGTV, PTV, and a 3 cm thick shell around PTV (PTV3cm). Other analyzed features were tumor sphericity (numeric index representing the regularity of the external surface of the tumor), CMD, PTV volume, and treatment technique.

Dose values were converted into BED (α/β = 10 Gy). Youden’s J statistic was applied for all analysed features to find the optimal cut-points [16]. The cut-point with the strongest DM-predictive power was determined as a receiver operating characteristic (ROC) threshold with the highest Youden’s index. Kaplan–Meier (KM) analysis and log-rank testing were employed to determine the significance of the optimal cut points [17].

Results

The DL-surv model c-index in cross-validation was 0.60 ± 0.02 and the testing c-index was 0.61. Figure 2 shows Grad-CAM heat maps for dose distributions of 9 randomly selected testing patients. One can see that the attention of the model is focused on the shell around the PTV.

Fig. 2
figure 2

Examples of dose distributions (9 different patients) with Grad-CAM heatmaps and PTV masks (pink color)

Full size image

The statistical significance of studied metrics as DM predictors in all patients and its IMRT/VMAT (sub)cohorts are summarized in Table 2. The optimal cut-points for all metrics are in Appendix C. The most significant predictors (p < 0.001) are clinical parameters (treatment technique or staging) and dose features in the shell around PTV, which confirms the grad-CAM results.

Table 2 Significances of DM-free predictors
Full size table

Since the treatment technique and several IGTV/PTV dose metrics are significant DM predictors, planning dose characteristics of IMRT and VMAT patients were compared in Appendix D. The chi-square test (χ2) comparing PTV Dmean distributions in IMRT and VMAT was p < 0.001, suggesting a highly significant difference between IMRT and VMAT planning dose, with VMAT delivering higher doses.

There is a strong correlation between some of the analyzed metrics. The strongest correlations are between Dmean in PTV3cm (Dmean,PTV3cm), PTVVol, sphericity, and treatment technique. Pearson’s correlation coefficients are ρ(Dmean,PTV3cm, PTVVol) = 0.62 (p < 0.001), ρ(Dmean,PTV3cm, sphericity) = − 0.53 (p < 0.001) and ρ(Dmean,PTV3cm, treat. technique) = − 0.54 (p < 0.001). The latter value indicates a lower Dmean,PTV3cm, for VMAT compared to IMRT. More details are in Appendix E.

Dmean in PTV3cm is statistically correlated with treatment technique, tumor sphericity, and PTVVol. Therefore, these metrics were additionally analyzed as DM predictors in patients with tumor sphericities > 0.5 (Table 3 and Appendix F). A sphericity threshold of 0.5 was used as it was the optimal cut-point. The group of patients with tumor sphericity > 0.5 found the treatment technique to be the only significant predictor (p = 0.02). Finally, the only significant predictor of DM in the “VMAT&Sphericity > 0.5” (sub)cohort was Dmean,PTV3cm with HR = 0.36 (95% CI: 0.17–0.79). Figure 3 shows Kaplan–Meier estimates for the optimal cut-point (Dmean,PTV3cm = 19.4 Gy BED).

Table 3 Significances of DM-free predictors
Full size table
Fig. 3
figure 3

Kaplan–Meier estimates and log-rank test results for the optimal cut-point of mean dose in PTV3cm for distant metastasis. The analysis was performed in a (sub)cohort of patients with tumor sphericity > 0.5 (more spherical) treated with VMAT

Full size image

Discussion

Our deep learning model allows localizing dose features with the strongest influence on DM. Previous machine learning models applied in the prediction of distant metastasis did not provide this, as they were extracting features only from CT or PET images [18,19,20], which limited their explainability. The limited predictive power of the model (c-index: 0.61) is likely due to the limited number of samples and the class imbalance of event rates (DM:DM-free = 1:4). Moreover, the strongest prognostic factors for distant metastasis are often related to tumor radiomic features [21], which might not be sufficiently captured in these images with a low spatial resolution of 3 × 3 × 3 mm3. However, this does not diminish the use of DL-surv in providing valuable inference insights into important dose features regarding DM.

As shown in Fig. 2, the model focuses almost exclusively on areas outside the PTV (pink color). Even though it is difficult to interpret the grad-CAMs to provide specific conclusions, PTV shape and the dose around it are highly important for the model decision-making, which agrees with previous studies [2,3,4].

As can be seen in Table 2, the strongest predictors of DM in all patients are treatment technique, tumor features, and dose features in PTV3cm (p < 0.001). Interestingly, the HR of Dmean in PTV3cm for the optimal cut point 26.2 Gy BED (Appendix C) is significantly larger than 1, indicating an increased risk of DM for higher doses outside PTV. Such a result conflicts with Diamant [2, 3], although it agrees with the findings of Lalonde [4], who used a dataset of the same mixed treatment techniques (IMRT/VMAT) as in our study. Lalonde showed a reduced incidence of DM in patients receiving lower Dmean in PTV3cm. This might be attributed to smaller target volumes, which tend to have more rapid dose fall-off, and various treatment techniques included in their dataset. In Diamant’s study, smaller tumors were associated with higher risk of DM compared to larger tumors, which contradicts previously published results [22,23,24,25]. Therefore, the associations between dose outside the PTV and DM, reported by Diamant, might likely be biased and related to some confounding factors which were not included in the analysis. Another confounding factor could be the treatment technique. Their dataset included patients who underwent 3D conformal radiotherapy (3D-CRT) or volumetric arc therapy (VMAT). The rate of DM was 25% in 3D-CRT patients vs 18% in VMAT patients. Diamant’s subsequent paper also claimed a lower incidence of DM in CK patients compared to 3D-CRT and VMAT patients [3]. However, a higher dose outside the PTV in CK also correlates with a higher dose in the PTV. The average prescribed dose to PTV in their CK dataset was higher compared to the rest of the patients (160 Gy of biologically effective dose [BED] vs 144 Gy BED), which might have introduced another bias concerning the results.

Dose features in PTV3cm are strongly correlated with tumor features (PTV volume, Sphericity) and treatment technique (see Appendix E). This supports the idea that HR > 1 is more related to tumor size and treatment technique than Dmean in PTV3cm. Smaller and more spherical tumors exhibit a lower risk of DM [22,23,24,25] and have a sharper dose fall-off (i.e., lower Dmean in PTV3cm) [26, 27]. Treatment technique is an important influencing factor as VMAT delivers more conformal doses into PTV [28, 29]. As is evident from Appendix D, the dose delivered into the PTV using VMAT was also significantly higher (pχ2 < 0.001) compared to IMRT, which might reduce the risk of distant metastasis. Moreover, VMAT provides a sharper dose fall-off (see Appendix D). Consequently, tumors treated with VMAT have a lower risk of DM while having lower Dmean in the PTV3cm. Another contributing factor might be dose interplay, which can lead to reduced dose to a mobile tumor treated with static IMRT compared to VMAT [30,31,32,33,34,35].

Since the treatment technique was found to be a significant prognostic factor (Tables 1, 2), we analyzed IMRT and VMAT (sub)cohorts independently. As is evident from Appendix E, there is no correlation between tumor features and treatment technique, implicating no bias caused by tumor characteristics in IMRT and VMAT (sub)cohorts. PTV volume, sphericity, and clinical maximum tumor diameter are not strong predictors in VMAT patients, although some are significant in IMRT patients (Table 2). This is presumably due to differences between IMRT and VMAT dose characteristics, see Appendix D. VMAT provides better PTV coverage even for larger tumors with less spherical shape. At the same time, IMRT may struggle to maintain the same coverage for larger and irregular tumors [26, 27]. As a result, the shape and size of the tumor are significant predictors of DM in IMRT patients. Also, Dmean in the IGTV is a significant predictor in IMRT patients and not in VMAT patients, which confirms the hypothesis that VMAT can keep the dose inside IGTV sufficiently high, no matter the tumor size or shape, while IMRT is more affected by tumor features. Over time, there has been a trend at our institution to drive the IGTV Dmax dose higher. The IMRT patients were treated longest ago, which might also explain some of this variance.

Table 1 also shows no significance of dose metrics in PTV3cm for DM prediction in VMAT (sub)cohorts. This confirms the results of Hughes [6], who included only VMAT patients and reported no significance of PTV3cm dose features.

Due to the prognostic power of tumor volume and sphericity for DM, a cohort restricted to tumor sphericities above 0.5 and its IMRT/VMAT (sub)cohorts were analyzed independently (see Table 3). The results show treatment technique and Dmean in the PTV as the only significant predictors of DM in patients with tumor sphericities above 0.5. Apparently, the treatment technique is again a significant predictor due to dose characteristics in IMRT and VMAT patients. VMAT patients in the analyzed dataset have higher doses in the PTV which presumably reduces the risk of DM. That is also why Dmean of the PTV is a significant predictor as well (p = 0.02). There were no significant prognostic parameters in the “IMRT&Sphericity > 0.5" (sub)cohort. However, the “VMAT&Sphericity > 0.5” (sub)cohort exhibits a high significance for Dmean in the PTV3cm. A higher dose in the 3 cm ring outside PTV is associated with a lower risk of DM. This confirms Diamant’s earlier findings [2, 3], even though there is a difference in our optimal cut points. Diamant reported an optimal cut point around 21 Gy EQD2 (~ 25 Gy BED), while our data indicates the optimal cut point as 19.4 Gy BED (see Fig. 3). This difference might result from deviations between his dataset and our VMAT&Sphericity > 0.5 (sub)cohort. In Diamant’s study [2], the median PTV volume in patients experiencing DM was 18.5 cm3 (5.7–51.5 cm3), while in our VMAT&Sphericity > 0.5 dataset it was 18.9 cm3 (7.9–40.3 cm3). The medians are comparable, although the range of PTV volumes in our data is concentrated around smaller values. Hence, the range of Dmean in PTV3cm is presumably also concentrated around lower values since smaller PTV volumes usually have sharper dose fall-off [26, 27]. Moreover, our dose grid resolution may have differed from Diamant’s, potentially contributing to additional minor discrepancies. Even though the optimal Dmean in the PTV3cm cut point is disputable and should be verified on a larger dataset, the trend of a lower rate of DM for larger doses around PTV might be correct for small spherical tumors treated with VMAT.

In conclusion, this study shows that the main reason for conflicting findings in previously published studies [2, 4, 6] was inconsistency in the datasets and insufficiently-considered confounding variables. This work reproduced the results of all previous publications by properly stratifying the patient dataset according to these variables. There is no causal correlation between the risk of DM and dose outside the PTV. However, the probability of DM decreases for higher doses outside the PTV in small spherical tumors treated with VMAT. The optimal cut point of the mean dose to PTV3cm for VMAT patients with tumor sphericities > 0.5 would be around 19–21 Gy BED. This would clinically imply considering larger PTV margins for smaller tumors (e.g., if IGTV > 2 cm, then margin ≤ 7 mm, else margin > 7 mm), although it requires verification on an independent dataset.

Availability of data and materials

Due to privacy and ethical concerns, the source data cannot be made available.

Abbreviations

3D-CRT:

3-Dimensional conformal radiotherapy

BED:

Biologically effective dose

bs:

Batch size

CK:

CyberKnife

CMD:

Clinical maximum tumor diameter

CNN:

Convolutional neural network

CT:

Computed tomography

CV:

Cross-validation

DM:

Distant metastasis

do:

Dropout

EQD2:

Equivalent dose in 2 Gy fractions

GRAD-CAM:

Gradient-weighted Class Activation Mapping

HP:

Hyperparameter

HR:

Hazard ratio

IGTV:

Internal gross tumor volume

IMRT:

Intensity-modulated radiation therapy

IQR:

Interquartile range

KM:

Kaplan–Meier

LR:

Local recurrence

lr:

Learning rate

LRR:

Locoregional recurrence

NOS:

Not otherwise specified

NSCLC:

Non-Small Cell Lung Cancer

OS:

Overall survival

PET:

Positron emission tomography

PTV:

Planning Target Volume

ROC:

Receiver operating characteristic curve

RT:

Radiotherapy

SBRT:

Stereotactic body radiotherapy

TPS:

Treatment planning system

VMAT:

Volumetric Modulated Arc Therapy

References

  1. Dyba T, Randi G, Bray F, Martos C, Giusti F, Nicholson N, et al. The European cancer burden in 2020: Incidence and mortality estimates for 40 countries and 25 major cancers. Eur J Cancer. 2021;157:308–47. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.ejca.2021.07.039.

    Article  PubMed  PubMed Central  Google Scholar 

  2. Diamant A, Chatterjee A, Faria S, El NI, Bahig H, Filion E, et al. Can dose outside the PTV influence the risk of distant metastases in stage I lung cancer patients treated with stereotactic body radiotherapy (SBRT)? Radiother Oncol. 2018;128:513–9. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.radonc.2018.05.012.

    Article  PubMed  Google Scholar 

  3. Diamant A, Heng VJ, Chatterjee A, Faria S, Bahig H, Filion E, et al. Comparing local control and distant metastasis in NSCLC patients between CyberKnife and conventional SBRT. Radiother Oncol. 2020;144:201–8. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.radonc.2020.01.017.

    Article  PubMed  Google Scholar 

  4. Lalonde R, Abdelhakiem M, Keller A, Huq MS. Dosimetric parameters related to occurrence of distant metastases and regional nodal relapse after SBRT for early-stage non-small cell lung cancer. Radiother Oncol. 2022;169:90–5. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.radonc.2022.02.019.

    Article  PubMed  Google Scholar 

  5. Yang K, Sun C, Yin R. Letter to the editor of radiotherapy and oncology regarding of the article “Dosimetric parameters related to occurrence of distant metastases and regional nodal relapse after SBRT for early-stage non-small cell lung cancer” by Lalonde et al. Radiother Oncol. 2022;177:240–1. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.radonc.2022.05.028.

    Article  CAS  PubMed  Google Scholar 

  6. Hughes RT, Steber CR, Jacobson TJ, Farris MK. Impact of dose to lung outside the planning target volume on distant metastasis or progression after SBRT for early-stage non-small cell lung cancer. Radiother Oncol. 2021;159:28–32. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.radonc.2021.03.004.

    Article  PubMed  PubMed Central  Google Scholar 

  7. Lalonde R, Abdelhakiem M, Keller A, Huq MS. Response to “Letter to the editor of radiotherapy and oncology regarding of the article ‘Dosimetric parameters related to occurrence of distant metastases and regional nodal relapse after SBRT for early-stage non-small cell lung cancer’ by Lalonde et al. Radiother Oncol. 2022;177:242. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.radonc.2022.05.032.

    Article  CAS  PubMed  Google Scholar 

  8. Cui S, Ten Haken RK, El Naqa I. Integrating multiomics information in deep learning architectures for joint actuarial outcome prediction in non-small cell lung cancer patients after radiation therapy. Int J Radiat Oncol Biol Phys. 2021;110:893–904. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.ijrobp.2021.01.042.

    Article  PubMed  PubMed Central  Google Scholar 

  9. Gensheimer MF, Narasimhan B. A scalable discrete-time survival model for neural networks. PeerJ. 2019;7: e6257. https://doiorg.publicaciones.saludcastillayleon.es/10.7717/peerj.6257.

    Article  PubMed  PubMed Central  Google Scholar 

  10. Borm KJ, Oechsner M, Wiegandt M, Hofmeister A, Combs SE, Duma MN. Moving targets in 4D-CTs versus MIP and AIP: comparison of patients data to phantom data. BMC Cancer. 2018. https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12885-018-4647-4.

    Article  PubMed  PubMed Central  Google Scholar 

  11. Dudas D, Dilling TJ, El Naqa I. Improved outcome models with denoising diffusion. Physica Med. 2024;119:103307. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.ejmp.2024.103307.

    Article  CAS  Google Scholar 

  12. Dudas D, Saghad PG, Dilling TJ, Perez BA, Rosenberg SA, El NI. Deep learning-guided dosimetry for mitigating local failure of non-small cell lung cancer patients receiving SBRT. Int J Radiat Oncol Biol Phys. 2024. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.ijrobp.2023.11.059.

    Article  PubMed  Google Scholar 

  13. Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD Statement. BMC Med. 2015. https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12916-014-0241-z.

    Article  PubMed  PubMed Central  Google Scholar 

  14. Uno H, Cai T, Pencina MJ, D’Agostino RB, Wei LJ. On the C-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data. Stat Med. 2011;30:1105–17. https://doiorg.publicaciones.saludcastillayleon.es/10.1002/sim.4154.

    Article  PubMed  PubMed Central  Google Scholar 

  15. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: visual explanations from deep networks via gradient-based localization. Int J Comput Vis. 2020;128:336–59. https://doiorg.publicaciones.saludcastillayleon.es/10.1007/s11263-019-01228-7.

    Article  Google Scholar 

  16. Kim S, Myong JP, Suh H, Lee KE, Youn Y-K. Optimal cutoff age for predicting mortality associated with differentiated thyroid cancer. PLoS ONE. 2015;10:e0130848. https://doiorg.publicaciones.saludcastillayleon.es/10.1371/journal.pone.0130848.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  17. Dudley N, Wickham RN, Aocn R, Coombs N. An introduction to survival statistics: Kaplan-Meier analysis. J Adv Pract Oncol. 2016. https://doiorg.publicaciones.saludcastillayleon.es/10.6004/jadpro.2016.7.1.8.

    Article  PubMed  Google Scholar 

  18. Gao H, He Z, Du X, Wang Z, Xiang L. Machine learning for the prediction of synchronous organ-specific metastasis in patients with lung cancer. Front Oncol. 2022;12:817372.

    Article  PubMed  PubMed Central  Google Scholar 

  19. Talebi A, Celis-Morales CA, Borumandnia N, Abbasi S, Pourhoseingholi MA, Akbari A, et al. Predicting metastasis in gastric cancer patients: machine learning-based approaches. Sci Rep. 2023;13:4163. https://doiorg.publicaciones.saludcastillayleon.es/10.1038/s41598-023-31272-w.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  20. Li Z, Wang S, Yu H, Zhu Y, Wu Q, Wang L, et al. A novel deep learning framework based mask-guided attention mechanism for distant metastasis prediction of lung cancer. IEEE Trans Emerg Top Comput Intell. 2023;7:330–41. https://doiorg.publicaciones.saludcastillayleon.es/10.1109/TETCI.2022.3171311.

    Article  Google Scholar 

  21. Yu L, Zhang Z, Yi H, Wang J, Li J, Wang X, et al. A PET/CT radiomics model for predicting distant metastasis in early-stage non–small cell lung cancer patients treated with stereotactic body radiotherapy: a multicentric study. Radiat Oncol. 2024;19:10. https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s13014-024-02402-z.

    Article  PubMed  PubMed Central  Google Scholar 

  22. Zhou H, Dong D, Chen B, Fang M, Cheng Y, Gan Y, et al. Diagnosis of distant metastasis of lung cancer: based on clinical and radiomic features. Transl Oncol. 2018;11:31–6. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.tranon.2017.10.010.

    Article  PubMed  Google Scholar 

  23. Wu J, Gensheimer MF, Zhang N, Han F, Liang R, Qian Y, et al. Integrating tumor and nodal imaging characteristics at baseline and mid-treatment computed tomography scans to predict distant metastasis in oropharyngeal cancer treated with concurrent chemoradiotherapy. Int J Radiat Oncol Biol Phys. 2019;104:942–52. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.ijrobp.2019.03.036.

    Article  PubMed  PubMed Central  Google Scholar 

  24. Coroller TP, Grossmann P, Hou Y, Rios Velazquez E, Leijenaar RTH, Hermann G, et al. CT-based radiomic signature predicts distant metastasis in lung adenocarcinoma. Radiother Oncol. 2015;114:345–50. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.radonc.2015.02.015.

    Article  PubMed  PubMed Central  Google Scholar 

  25. Fried DV, Tucker SL, Zhou S, Liao Z, Mawlawi O, Ibbott G, et al. Prognostic value and reproducibility of pretreatment CT texture features in stage iii non-small cell lung cancer. Int J Radiat Oncol Biol Phys. 2014;90:834–42. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.ijrobp.2014.07.020.

    Article  PubMed  PubMed Central  Google Scholar 

  26. Brito Delgado A, Cohen D, Eng TY, Stanley DN, Shi Z, Charlton M, et al. Modeling the target dose fall-off in IMRT and VMAT planning techniques for cervical SBRT. Med Dosim. 2018;43:1–10. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.meddos.2017.07.009.

    Article  CAS  PubMed  Google Scholar 

  27. Narayanasamy G, Desai D, Maraboyina S, Penagaricano J, Zwicker R, Johnson E. A dose falloff gradient study in RapidArc planning of lung stereotactic body radiation therapy. J Med Phys. 2018;43:147–54. https://doiorg.publicaciones.saludcastillayleon.es/10.4103/jmp.JMP_38_18.

    Article  PubMed  PubMed Central  Google Scholar 

  28. Li Y, Wang J, Tan L, Hui B, Ma X, Yan Y, et al. Dosimetric comparison between IMRT and VMAT inirradiation for peripheral and central lung cancer. Oncol Lett. 2018;15:3735–45. https://doiorg.publicaciones.saludcastillayleon.es/10.3892/ol.2018.7732.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  29. Xu Y, Deng W, Yang S, Li P, Kong Y, Tian Y, et al. Dosimetric comparison of the helical tomotherapy, volumetric-modulated arc therapy and fixed-field intensity-modulated radiotherapy for stage IIB-IIIB non-small cell lung cancer. Sci Rep. 2017;7:14863. https://doiorg.publicaciones.saludcastillayleon.es/10.1038/s41598-017-14629-w.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  30. Leste J, Medjahed I, Arnaud F-X, Ferrand R, Franceries X, Bardies M, et al. A study of the interplay effect for VMAT SBRT using a four-axes motion phantom. J Appl Clin Med Phys. 2020;21:208–15. https://doiorg.publicaciones.saludcastillayleon.es/10.1002/acm2.12947.

    Article  PubMed  PubMed Central  Google Scholar 

  31. Stambaugh C, Nelms BE, Dilling T, Stevens C, Latifi K, Zhang G, et al. Experimentally studied dynamic dose interplay does not meaningfully affect target dose in VMAT SBRT lung treatments. Med Phys. 2013;40:091710. https://doiorg.publicaciones.saludcastillayleon.es/10.1118/1.4818255.

    Article  PubMed  Google Scholar 

  32. Ong CL, Dahele M, Slotman BJ, Verbakel WFAR. Dosimetric impact of the interplay effect during stereotactic lung radiation therapy delivery using flattening filter-free beams and volumetric modulated arc therapy. Int J Radiat Oncol Biol Phys. 2013;86:743–8. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.ijrobp.2013.03.038.

    Article  PubMed  Google Scholar 

  33. Jensen MD, Abdellatif A, Chen J, Wong E. Study of the IMRT interplay effect using a 4DCT Monte Carlo dose calculation. Phys Med Biol. 2012;57:N89. https://doiorg.publicaciones.saludcastillayleon.es/10.1088/0031-9155/57/8/N89.

    Article  PubMed  Google Scholar 

  34. Zou W, Yin L, Shen J, Corradetti MN, Kirk M, Munbodh R, et al. Dynamic simulation of motion effects in IMAT lung SBRT. Radiat Oncol. 2014;9:225. https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s13014-014-0225-3.

    Article  PubMed  PubMed Central  Google Scholar 

  35. Watkins W, Song WY, Hugo GD, Weiss E, Siebers JV. The effects of interplay on accumulated dose in high-dose-rate stereotactic body radiation therapy of lung cancer. Int J Radiat Oncol Biol Phys. 2013;87:S758. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.ijrobp.2013.06.2009.

    Article  Google Scholar 

Download references

Acknowledgements

None.

Funding

This work was supported by National Institute of Health (NIH) grant R01-CA233487 and Department of Defense (DoD) Congressional Directed Med Res Prog (CDMRP) W81XWH-22-1-0277.

Author information

Authors and Affiliations

  1. Faculty of Nuclear Sciences and Physical Engineering, Czech Technical University in Prague, Prague, Czechia

    D. Dudas

  2. Department of Machine Learning, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, USA

    I. El Naqa

  3. Department of Radiation Oncology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, USA

    T. J. Dilling

Authors
  1. D. Dudas
    View author publications

    You can also search for this author inPubMed Google Scholar

  2. T. J. Dilling
    View author publications

    You can also search for this author inPubMed Google Scholar

  3. I. El Naqa
    View author publications

    You can also search for this author inPubMed Google Scholar

Contributions

DD: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data Curation, Writing—Original Draft, Visualization. TJD: Data Curation, Resources, Writing—Review & Editing, Supervision. IEN: Conceptualization, Methodology, Supervision, Project administration, Funding acquisition.

Corresponding author

Correspondence to D. Dudas.

Ethics declarations

Ethics approval and consent to participate

This study utilized Good Clinical Practice Guidelines, in accordance with the Belmont Report, and was approved by the Institutional Review Board at Moffitt Cancer Center (MCC 18883).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Dudas, D., Dilling, T.J. & Naqa, I.E. A deep learning-informed interpretation of why and when dose metrics outside the PTV can affect the risk of distant metastasis in SBRT NSCLC patients. Radiat Oncol 19, 127 (2024). https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s13014-024-02519-1

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s13014-024-02519-1

Radiation Oncology

ISSN: 1748-717X

Contact us