Pretreatment Spatially Aware Magnetic Resonance Imaging Radiomics Can Predict Distant Brain Metastases (DBMs) After Stereotactic Radiosurgery/Radiation Therapy (SRS/SRT)

Abstract

Purpose

Stereotactic radiosurgery/radiation therapy (SRS/SRT) increasingly has been used to treat brain metastases. However, the development of distant brain metastases (DBMs) in the untreated brain remains a serious complication. We sought to develop a spatially aware radiomic signature to model the time-to-DBM development in a cohort of patients leveraging pretreatment magnetic resonance imaging (MRI) and radiation therapy treatment planning data including radiation dose distribution maps.

Methods and Materials

We retrospectively analyzed a cohort of 105 patients with brain metastases treated by SRS/SRT with pretreatment multiparametric MRI (T1, T1 postcontrast, T2, fluid-attenuated inversion recovery). Three-dimensional radiomic features were extracted from each MRI sequence within 5 isodose regions of interest (ROIs) identified via radiation dose distribution maps and gross target volume (GTV) contours. Clinical features including patient performance status, number of lesions treated, tumor volume, and tumor stage were collected to serve as a baseline for comparison. Cox proportional hazards (CPH) modeling and Kaplan–Meier analysis were used to model time-to-DBM development.

Results

CPH models trained using radiomic features achieved a mean concordance index (c-index) of 0.63 (standard deviation [SD], 0.08) compared with a c-index of 0.49 (SD, 0.09) for CPH models trained using clinical factors. A CPH model trained using both radiomic and clinical features achieved a c-index of 0.69 (SD, 0.08). The identified radiomic signature was able to stratify patients into distinct risk groups with statistically significant differences (P = .00007) in time-to-DBM development as measured by log-rank test. Clinical features were unable to do the same. Radiomic features from the peritumoral 50% to 75% isodose ROI and GTV region were most predictive of DBM development.

Conclusions

Our results suggest that radiomic features extracted from pretreatment MRI and multiple isodose ROIs can model time-to-DBM development in patients receiving SRS/SRT for brain metastases, outperforming clinical feature baselines. Notably, we believe we are the first to leverage SRS/SRT dose maps for ROI identification and subsequent radiomic analysis of peritumoral and untargeted brain regions using multiparametric MRI. We observed that the peritumoral environment may be implicated in DBM development for SRS/SRT-treated brain metastases. Our preliminary results might enable the identification of patients with predisposition to DBM development and prompt subsequent changes in disease management.

Introduction

Brain metastases (BMs) are a common occurrence among patients with advanced cancer, with the potential to cause significant morbidity. They can arise due to limited intracranial penetration of antineoplastic agents, as well as an immune-privileged status preventing synergy with the systemic immune response. BM incidence will likely continue to increase as survival among patients with metastases, particularly those with oligometastases, increases; the latter trend is due at least in part to advancements in systemic agents like immunotherapy¹ and more effective local consolidation.² BM-directed intracranial therapy, such as stereotactic radiosurgery (SRS)/stereotactic radiation therapy (SRT) and/or craniotomy, has thus become increasingly important in the management of BMs, displacing older treatments, including whole-brain radiation therapy. However, in favoring more localized approaches, the development of distant brain metastases (DBMs) outside of the targeted area remains elevated at ∼50% or greater at 12 months’ posttreatment,3, 4, 5 even in patients on maximal systemic therapy.

Knowing the true risk of DBM development and ideally the associated time–course would allow providers additional tools for clinical decision-making. The most practical use of such information would be a change in the surveillance interval after SRS/SRT, with a greater-risk patient being imaged more frequently. In specific cases in which a practitioner feels either SRS/SRT or whole-brain radiation therapy may be reasonable (perhaps a patient with 10+ lesions), such insight may provide further justification for choosing a specific strategy. The question of how many lesions are suitable for SRS/SRT continues to be perplexing. It may be the case that the true risk of DBMs is driven by biology, which in turn manifests as characteristic imaging phenotypes, rather than simply the number of lesions at presentation. Clinical features, such as performance status, age, or number of BMs can indeed be prognostic^6,7 and are incorporated into some predictive models.⁸ There are further some standardized guidelines for assessing treatment response radiographically.⁹ However, neither of these approaches provide granular patient-level insight into DBM development.

Radiomics is a fast-growing method of radiology image analysis that enables the extraction of quantitative measurements of structure, intensity, morphology, and heterogeneity that are imperceptible to the human eye. In certain tumors, these subvisual features have been shown to predict tumor response to radiation therapy (RT),¹⁰ chemotherapy,¹¹ and immunotherapy.¹² In this retrospective study, we hypothesized that magnetic resonance imaging (MRI)-based radiomics would provide clinically actionable insight into the risk for DBMs that would outperform canonical clinical characteristics. Since contrast-enhanced MRI is a standard modality used for SRS/SRT planning, as well as posttreatment surveillance, it is an attractive target for radiomic analysis. Indeed, the role of MRI-based radiomic models in prediction of response to SRS/SRT has been reported.13, 14, 15, 16, 17 However, these previous works have largely studied local control as measured by changes in tumor volume after SRS/SRT and primarily investigated methods for binary classification of patient response to therapy.

Here, we attempt to predict the development of DBMs rather than local tumor control and leverage regression analysis to more accurately model the timeline of disease progression. A novel addition in this study is the incorporation of isodose-based regions of interest (ROIs), which were selected to represent the tumor volume/tumor bed, the peritumoral area or tumor microenvironment (TME), and untreated brain. Areas outside of the tumor itself might harbor important predictive information, as demonstrated in studies of the TME and its associated prognostic value.^11,15,18,19 Although there are indeed several ways to select ROIs, our approach offered some advantages, in our view. First, it was simple to implement, as the plan data were readily available, obviating the need for additional auto- or manual segmentation. Second, as plans were ultimately approved by the treating radiation oncologist, these regions would naturally incorporate some anatomic context (dose fall-offs would avoid critical organs such as the brain stem for instance). Third, in our data set the dose distributions would be consistent between patients given the similar treatment planning methodology.

By studying radiomic features from multiple isodose ROIs, we hope to model some of these processes to improve predictive capability and provide interpretable insights into the regions most implicated in DBM predisposition. An overall pipeline for this study is contained in Fig. 1.

Figure 1.

Study pipeline. (a) visualizes the MRI and RT dose map data studied in addition to the processing applied to identify isodose ROIs. (b) highlights the radiomic feature extraction framework in which radiomic features are calculated from multiple MRI sequences and isodose ROIs. (c) shows the machine learning experiments performed including analysis of clinical features as a baseline comparison, identification of informative radiomic features, and time-to-event modeling of DBM development. Abbreviations: DBM = distant brain metastasis; MRI = magnetic resonance imaging; ROI = region of interest; RT = radiation therapy. (A color version of this figure is available at 10.1016/j.adro.2024.101457.)

Methods and Materials

Data set

In this institutional review board–approved single-institution retrospective study (IRB2020-00502), our treatment planning system, ARIA from Varian Medical Systems,²⁰ was queried for all treatment courses delivered to the brain between January 1, 2014, and March 1, 2021. These were filtered (Fig. 2) to select our final cohort of 105 patients. Patients’ charts were manually assessed for DBMs, defined as tumor recurrence in the brain outside of the treated target volumes. Computed tomography (CT) simulation, structure sets, planning data (RT-Digital Imaging and Communications in Medicine), as well as pretreatment MRI including T1, contrast-enhanced T1 (T1-weighted), T2, and fluid-attenuated inversion recovery (FLAIR) sequences, were exported for this group. The structure sets housed contours of organs at risk including the brain as well as gross tumor volumes (GTVs), clinical tumor volumes, and planning tumor volumes. Planning data contained the deposited dose in gray (Gy) delivered to each voxel of the associated planning CT. All exported imaging and RT planning data used in the analysis were originally approved by the treating radiation oncologist and were identical to those used during treatment delivery. To improve the generalizability of the trained models, SRS/SRT courses for a single patient were treated as independent if they occurred more than 60 days apart. Models were trained on MRI data taken before each SRS treatment course and evaluated for DBM occurrence after individual courses.

Figure 2.

Patient selection. Patients were selected on the basis of receiving SRS/SRT with pretreatment MRI and posttreatment imaging to confirm the incidence of DBM. Abbreviations: DBM = distant brain metastasis; MRI = magnetic resonance imaging; SRS, = stereotactic radiosurgery; SRT = stereotactic radiation therapy.

Preprocessing and isodose ROI delineation

MRI scans were automatically and rigidly coregistered to CT simulation scans using a custom software pipeline. Registrations were also verified by external review. For each MRI scan, the skull was removed using the physician-approved brain contour contained in the corresponding structure set. The remaining brain image was then normalized via z score (voxels with intensities lower than the 0.01 percentile and higher than the 99.9 percentile were clipped before normalization). Finally, images were resampled to a 1-mm × 1-mm × 1-mm isotropic volume. Skull stripping and isotropic resampling were also applied to the RT dose maps.

RT dose maps were leveraged to automatically select ROIs for radiomic feature analysis. Regions receiving 0% to <25%, 25% to <50%, 50% to <75%, and 75% of the dose were selected representing the untreated brain, regional brain, immediate peritumoral region, and tumor volume/tumor bed, respectively (Fig. 1a). In addition, radiomic features were extracted from the physician-approved GTV structure, yielding a total of 5 ROIs for feature extraction from 4 MRI sequences (Fig. 1b).

Radiomic feature extraction

Three-dimensional radiomic features were extracted using an in-house MATLAB 2021a (MathWorks, Natick, MA) pipeline. Radiomic features studied included gradient, Haralick,²¹ Laws energy,²² Gabor wavelet,²³ and CoLlAGe²⁴ features. Gradient features measure the rate of voxel intensity changes in an image. Haralick features are calculated from the gray-level co-occurrence matrix and measure heterogeneity in image texture. Laws energy features calculate the prevalence of edges, spots, ripples, and waves at different resolutions within an image. Gabor wavelet features measure image patterns at different frequencies and orientations. CoLlAGe features quantify heterogeneity in intensity gradients across the image. Each of these features has been previously studied and shown to model pathologic and physiological processes in various oncologic diagnostic and prognostic applications.^18,24, 25, 26

In total, 316 radiomic features were extracted from each of the 5 ROIs and 4 MRI sequences, yielding a total of 7320 features per patient. To understand the contribution of features from each MRI sequence and ROI combination, experiments were performed using individual subsets as well as the total pooled feature set. Radiomic features were initially calculated for each voxel within the delineated ROI. These voxel-wise values were used to generate descriptive statistics of feature expression distributions including measures of mean, median, variance, kurtosis, and skewness. In addition to radiomic features, clinical features such as Eastern Cooperative Oncology Group performance status, target volume (determined from GTV/clinical tumor volume contours), number of metastases, primary tumor staging, and patient age were collected for each patient to serve as a baseline comparison.

Radiomic feature selection

Because of the large feature space described previously, we presumed many of the extracted features would contain redundant or irrelevant information that can result in overfitting or ineffective model generalization. Thus, feature selection was performed to train machine-learning models on the most predictive radiomic features. First, we filtered our total radiomic feature set to remove highly correlated features. Specifically, if multiple features demonstrated a Pearson correlation coefficient of 0.9 or higher, only one was retained. Then, univariate Cox proportional hazards (CPH) models²⁷ were trained to model the time course for DBM development using each individual radiomic and clinical feature independently. Statistical significance was determined for each model using the log-likelihood rank test. Univariate models found to be most statistically significant were selected and their features used for multivariate model construction.

Time-to-DBM analysis

CPH models have commonly been used to model time-to-event data including for patient survival^28,29 and patient response to treatment.^16,28 Twenty multivariate CPH models were trained on top-performing radiomic features (selected as described previously) obtained from within each of the 20 MRI-ROI combinations. Multivariate CPH models were then trained from top performing radiomic features from all isodose ROIs, all MRI sequences, and the entire set of radiomic features studied. A multivariate CPH model was also trained for top performing clinical features.

Hazard ratios from multivariate CPH models is given by $h_x=e^{{\beta }_x}$ where $h_x$ denotes the hazard ratio for a feature $x$ and ${\beta }_x$ denote the CPH model coefficient for $x.$ A risk score for patients was calculated as $R_p=\sum _i{e}^{x_i{\beta }_{x_i}}$, where $R_p$ denotes an individual patient's risk score calculated for $i$ features using feature expression values ($x_i$) and corresponding CPH model weights (${\beta }_{x_i}$). The median risk score was used as a threshold for patient stratification. Patients with a score greater than median were assigned to the high-risk group, and patients with submedian scores were assigned to the low-risk group. Both a radiomic risk score and clinical risk score were created. Kaplan–Meier curves were used to analyze time-to-DBM development in high and low-risk patient groups identified by each risk score.

Statistical analysis

All experiments were performed in a 5-fold cross validation setting to ensure model robustness and generalizability. Folds were created in a patient-wise manner such that all treatment plans from an individual patient could only appear in either the training or testing set for each iteration of cross-validation. This served as an additional guard against bias that might arise from analysis of multiple treatment plans for a single patient. Training folds were used for univariate CPH model feature selection and multivariate model training. Univariate models were compared by P value obtained using the log-likelihood rank test. Trained multivariate models were evaluated by concordance index (c-index) on the test fold. The c-index is a measure of a regression model's ability to correctly model the order of individual patient's survival times; a score of 1 indicates perfect accuracy and a score of 0.5 indicates performance equal to random chance. Median risk score thresholds were determined on the training folds and used to stratify the test fold into low and high-risk groups. Throughout a full iteration of cross-validation, each patient was assigned to a low or high-risk group when a part of the test fold. Statistical significance for risk stratification was evaluated using the log rank test for time-to-DBM development differences among the low and high-risk groups identified. To correct for the multiple comparisons being made for different ROI and MRI sequence combinations in addition to clinical and radiomic model comparisons, the Bonferroni correction method was employed. A total of 30 MRI sequence and ROI combinations were tested in addition to comparisons with clinical and combined radiomic and clinical models. Using an initial alpha of 0.01 and this summed total of 32 comparisons resulted in an adjusted alpha of 0.00031 after Bonferroni correction.

Results

Patient characteristics

In total, 755 patients were identified from our institution receiving RT to the brain either definitively, adjuvantly, or both (treatment to a resected cavity and unresected lesions simultaneously) between 2014 and 2021. Of these, 402 patients received SRS/SRT, as defined by a dose/fraction >5 Gy. Then, 239 patients were selected with an available pretreatment MRI before SRS/SRT. After manual curation, that is, removing patients with diagnoses other than metastatic solid tumors and/or limited follow-up (no posttreatment MRI), the final cohort was 105 patients treated with 126 courses of SRS/SRT. Demographic information is available in Table 1. Information regarding SRS/SRT treatment and subsequent DBM development can be found in Table E3.

Table 1.

Patient cohort

Age, y	63.5 (24.8-86.3)	N stage	N = 67
Number of lesions	1.4 (1-15)	0	23 (34.3%)
ECOG	N = 78	1	8 (11.9%)
0	9 (11.5%)	2	28 (41.8%)
1	38 (48.7%)	3	8 (11.9%)
2	19 (24.3%)	Primary tumor	N = 105
3	11 (14.1%)	Lung	58 (55.2%)
4	1 (1.3%)	Breast	19 (18.1%)
T stage	N = 68	Skin	7 (6.7%)
1	20 (29.4%)	Kidney	4 (3.8%)
2	21 (30.9%)	Unknown	3 (2.9%)
3	17 (25.0%)	Ovary	2 (1.9%)
4	10 (14.7%)	Thyroid	2 (1.9%)
		Other	10 (9.5%)

Abbreviations: ECOG = Eastern Cooperative Oncology Group.

Among these 105 patients, single-fraction doses ranged from 14 to 20 Gy (median, 18 Gy), whereas SRT regimens ranged from 22.5 to 40 Gy (median, 30 Gy) in 3 to 5 fractions (median, 5). All doses were prescribed to 100% isodose line. The median follow-up time was 355 days.

In total, 53 (50.5%) of 105 patients (67 of 126 courses) were found to have developed DBMs. The median time to DBM development was 118 days. No significant difference was observed in time-to-DBM development for patients receiving SRS and SRT (Table E3).

Time-to-DBM analysis

Results for multivariate CPH models trained from radiomic and clinical data are shown in Table 2. Multivariate CPH models trained independently on features from each MRI sequence achieved average c-indices between 0.43 and 0.63 (Table 2a). Most CPH models trained on radiomic features from individual MRI sequence and isodose ROI combinations resulted in c-indices near 0.50. CPH models trained on selected radiomic features from the union of all MRI modalities and isodose ROIs resulted in an average c-index of 0.63 (standard deviation, 0.08). The highest individual CPH model performance was observed with radiomic features from the 50% to 75% isodose region on the T1 MRI sequence. The addition of clinical features to radiomic features resulted in a multivariate CPH model with a mean c-index of 0.69 (standard deviation, 0.08). A multivariate CPH model trained using clinical features achieved an average c-index of 0.49 (standard deviation, 0.09) on 5-fold cross-validation (Table 2b).

Table 2.

Time-to-DBM C-index results

		Regions of interest
a.		0-<25%	25%-<50%	50%-<75%	75%	GTV	Whole brain
MRI sequence	T1Pre	0.54 ± 0.06	0.45 ± 0.09	0.63 ± 0.06	0.47 ± 0.09	0.55 ± 0.10	0.62 ± 0.06
	T1Post	0.52 ± 0.05	0.45 ± 0.06	0.50 ± 0.06	0.49 ± 0.08	0.58 ± 0.11	0.55 ± 0.15
	T2	0.45 ± 0.07	0.39 ± 0.07	0.42 ± 0.06	0.43 ± 0.11	0.45 ± 0.02	0.41 ± 0.09
	FLAIR	0.51 ± 0.04	0.48 ± 0.07	0.44 ± 0.03	0.48 ± 0.16	0.53 ± 0.02	0.53 ± 0.07
	All	0.51 ± 0.09	0.50 ±.08	0.57 ± 0.11	0.44 ± 0.11	0.58 ± 0.05	0.63 ± 0.08
b.	c-index
Clinical	0.49 ± 0.09
Radiomics	0.63 ± 0.08
Radiomics + clinical	0.69 ± 0.08

Abbreviations: FLAIR = fluid-attenuated inversion recovery; GTV = gross tumor volume; MRI = magnetic resonance imaging.

Bolded values indicate highest predictive performance as measured by c-index.

Kaplan–Meier curves visualizing time-to-DBM development for low- and high-risk patient groups created using radiomic features from various MRI sequences and ROI combinations are shown in Fig. 3. After Bonferroni correction for multiple comparisons, statistically significant separation of patients into low- and high-risk groups was observed using radiomic features from the T1 MRI sequence and 50% to 75% isodose ROI (P = .00015). A radiomic model trained on radiomic features from all isodose ROIs using only the T1 MRI sequence also achieved statistically significant risk stratification (P = .00022). Finally, the best performance in risk stratification was observed when radiomic features from all MRI sequences and isodose ROIs were considered for risk score creation (P = .00007). The addition of clinical features to radiomic features did not improve risk stratification for DBM development. Patients stratified into low and high-risk groups using clinical features did not have statistically significant differences in time-to-DBM development upon Kaplan–Meier curve analysis (P = .98) (Fig. 3b).

Figure 3.

Kaplan–Meier (KM) curves for DBM development. Visualized are KM curves showing low- and high-risk patient groups’ time-to-DBM development. (a) shows risk groups created using multivariate CPH models trained on radiomic features from various MRI sequence and ROI combinations. KM curves located in the “All ROIs” column and “All Sequences” row are created using CPH models trained on the union of the most predictive radiomic features extracted from within individual ROI-MRI sequence combinations. One asterisk and a surrounding red box denote statistical significance at a Bonferroni corrected alpha of 0.00031. The T1Pre-MRI sequence as well as the GTV and 50% to 75% isodose ROIs yielded radiomic features resulting in statistically significant stratification of risk for DBM development. (b) displays the best-performing radiomics risk stratification model using features from all ROIs and all MRI sequences in addition to a clinical baseline stratification model and a combined radiomic and clinical model. Abbreviations: CPH = Cox proportional hazards; DBM = distant brain metastasis; GTV = gross tumor volume; MRI = magnetic resonance imaging; ROI = region of interest; RT = radiation therapy. (A color version of this figure is available at 10.1016/j.adro.2024.101457.)

Top-performing features and associated hazard ratios used to generate patient risk stratification are shown in Fig. 4. The Laws E3E3S3 radiomic feature quantifying edges and spots extracted from the GTV ROI and FLAIR MRI was determined to be a statistically significant protective predictive factor (P = .0002) for DBM development. The Haralick correlation radiomic feature in the 0% to 25% isodose ROI on T1 MRI was a risk factor for DBM development (P = .016) and measures linear dependencies in the gray-level co-occurrence matrix. The Haralick information radiomic feature in the GTV ROI and T2 MRI was a protective predictive factor for DBM development (P = .038) and quantifies correlations between gray-level co-occurrence values, potentially indicating underlying patterns in texture within the GTV. Hazard ratios for clinical models were all determined to be statistically insignificant (Fig. 4). Figure 4b provides qualitative visualizations of the radiomic features used for patient DBM development modeling from a patient who developed DBMs and one who did not. Some intuitive correlations can be observed from these results. For instance, the patient who was negative for DBM exhibited greater expression values of the Haralick difference entropy feature in the 50% to 75% isodose on T1 MRI compared with the patient who was positive for DBM, aligning with the observed negative hazard ratio of 0.430. Similarly, greater GTV expression of the Laws E3S3S3 feature on FLAIR MRI is observed in the patient who was negative for DBM, with an associated hazard ratio of 0.412.

Figure 4.

Radiomic risk score features. (a) Displays a forest plot containing hazard ratios and confidence intervals for the best-performing radiomic risk score on cross-validation. Features with a hazard ratio less than one indicate a protective predictive effect whereas those with a hazard ratio greater than one indicate a risk factor. (b) visualizes radiomic feature expression for the 4 features in (a) with statistically significant or nearly statistically significant hazard ratios. The first row represents a patient who did not develop DBM, and the second row shows a patient who developed DBM. Abbreviations: CI = confidence interval; DBM = distant brain metastasis; FLAIR = fluid-attenuated inversion recovery; GTV = gross tumor volume; HR = hazard ratio. (A color version of this figure is available at 10.1016/j.adro.2024.101457.)

Discussion

In this work, we demonstrated that MRI-based radiomic features extracted from specific ROIs determined by radiation dose distribution can predict DBM development and, more significantly, its likely time course. Our model outperforms a similar approach using canonical clinical features in all experimental settings.

Previous works studying BMs that leverage imaging data do not generally attempt to study features from multiple compartments (such as the TME and/or untreated brain), and instead limit their focus to the metastatic lesions themselves.^13,14,16 Our work allows analysis from multiple ROIs identified automatically via isodose regions, which could be considered a form of “physician-guided” attention in the analysis pipeline. Among our models, the highest performance by c-index was observed with radiomic features from the 50% to 75% isodose ROI of T1 MRIs. This suggests that the TME, rather than the tumor volume itself, may provide more actionable insight into DBM development. Similar findings have previously been reported in the literature for both primary brain tumors such as glioblastoma multiforme¹⁸ and BMs.¹⁵

Specifically from our analysis, Haralick difference entropy and Haralick sum entropy expression in the 50% to 75% isodose ROI measure heterogeneity in image texture. Such features may have biologic correlates, such as tumor-infiltrating lymphocyte activity,¹¹ patterns in tumor cell distribution within the TME,¹⁵ or vasculature changes caused by malignant lesions.³⁰ Our model also suggests that features from the untreated brain may play a role in patient susceptibility to DBM development as demonstrated by high Haralick correlation signatures in the 0% to 25% isodose ROI of T1 MRIs (Fig. 4). This finding may represent some underlying homogeneity or close-knit structure within the untreated brain that predisposes a patient to DBM development.

Another advantage of isodose-based ROIs is that they are generated from dose maps already created during the SRS/SRT course, eliminating the need for computationally intense autosegmentation or preprocessing. It is conceivable that integrating this type of analysis into a treatment planning system should be relatively straightforward, and such an option would provide radiation oncologists key clinical insight at the time of planning.

Modeling the time course of DBM occurrence using radiomic features is a capability that, to our knowledge, has not been extensively explored. This information may be particularly important in clinical decision-making, as a likely “early” failure can preferentially be acted upon with more aggressive therapy and/or strict surveillance. Our data also confirm a great degree of variability in time-to-DBM development, as a tangible portion of patients may develop DBMs over a year after receiving initial SRS/SRT. For comparison, a nomogram developed by Ayala-Peacock et al⁸ also modeled the time course of DBM development, achieving a c-index of 0.63 when using clinical features alone. In our study, radiomic features from all MRI modalities and ROIs yielded a comparable c-index of 0.63; however, the combination of radiomic and clinical features resulted in a model with an improved c-index of 0.69. This improvement implies we are able to leverage complementary information contained in these variables to improve radiomic feature-based response prediction. It is further notable that the number of metastatic lesions present did not show a statistically significant association with DBM development in our cohort. This result can be contrasted with previous studies^7,8 but may suggest that specific patients with otherwise-significant disease burden can still be offered SRS/SRT if they are deemed to be low risk by our radiomic risk score.

Our study indeed has several areas of improvement. The primary shortcoming is the relatively small patient cohort studied (N = 105) and the heterogeneous mix of solid tumor histologies. Although this sample size is similar to other related works,13, 14, 15, 16 it is not feasible in our cohort to develop specific radiomic markers for underlying tumor/molecular subtypes. For instance, melanoma has been reported previously as a predictor of DBM development.^7,8 However, the inclusion of only 7 patients with a primary cancer of the skin in our study precluded a thorough analysis as to whether there might be a unique radiomic signature in this subset. We have performed a basic subset analysis demonstrating the performance of our signature on patients with breast and lung primary cancer due to their larger representation in our data set (Fig. E1), and we found that our radiomic risk score was able to stratify patients with lung cancer into low- and high-risk for DBM with statistical significance (P = .00049) but not patients with breast cancer (P = .1518). This may be due to the larger prevalence of patients with primary lung cancer in our data set, although we again note that our risk score is able to stratify the complete set of patients studied with statistical significance (P = .00007). Additionally, we determined to include repeat treatments for patients receiving more than a single course of SRS/SRT, designing our experiments to ensure no bias when accounting for repeat treatments. To validate this approach, we also demonstrate that our highest-performing radiomic signature can stratify patients into low and high risk for DBM groups using only their first course of SRS/SRT (P = .0004), in Figure E2. Furthermore, the retrospective nature of our study prevented the use of a standardized imaging protocol, resulting in heterogeneity in MRI scans acquired including differing scanning machines and imaging parameters. We have included a summary of these parameters in the supplementary materials (Table E1 and Table E2) in addition to experiments demonstrating radiomic feature expression stability within some of these variables (Fig. E3 and Fig. E4). We report parameters similar to those reported in other studies and have performed preprocessing steps (described in the Methods section) similar to those used in most radiomic analyses of brain MRI scans.13, 14, 15, 16^,31 As molecular profiling and mutation targeted therapies become increasingly prevalent in cancer treatment, our work might benefit from future experiments studying the interaction between predictive radiomic features and cancer-specific biomarkers, something that could not be thoroughly explored here due to inadequate data for many primary tumors. In the future, larger data sets composed of more homogenous patient populations might enable discovery of individualized radiomic signatures for BMs secondary to various primary cancers and indeed subtypes within those cancers such as ER+ versus ER– breast tumors. It is certainly feasible that the individual biology and driving mutations of different cancers might result in different rates of intracranial control after SRS/SRT and that radiomic signatures created for these individual disease types might outperform a generalized radiomic risk score. Nonetheless, the dearth of sufficient homogenous data makes doing so difficult, and we believe that our model may capture more generalized predictive features that enable some degree of risk stratification and time-to-event modeling of DBM development for patients with BMs with varied primary tumor types. Similarly, the limited cohort size prompted us to combine definitive and adjuvant treatments together in our analysis. These patients may indeed have divergent radiomic profiles which cannot be captured by our current results. Finally, our model currently lacks any incorporation of posttreatment imaging, which likely has important information that can further refine our predictions. We aim to explore posttreatment radiomic features in a delta-radiomic^32,33 setting in subsequent versions of this work.

Conclusions

Here we have developed an interpretable radiomic model for time-to-DBM development modeling based upon the use of isodose ROIs identified from radiation dose distribution maps. Our approach uniquely provides insight into the specific regions and MRI sequences predictive of DBM development while also modeling the time course of DBM development for specific patients. Additionally, because MRI and RT dose maps are routinely included in standard SRS/SRT treatment planning, our analysis can easily be integrated into current clinical workflows. In future work we anticipate that the inclusion of deep-learning modeling, particularly leveraging self-supervised approaches, might enable further analysis of both imaging and RT planning data including radiation dose maps for outcome prediction tasks.

Disclosures

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.