Voxel-wise prediction of recurrent high grade glioma via proximity estimation coupled multi-dimensional SVM

Abstract

Purposes:

To provide early and localized glioblastoma (GBM) recurrence prediction, we introduce a novel post-surgery multi-parametric MR-based support vector machine (SVM) method coupling with stem cell niches (SCN) proximity estimation.

Methods:

This study utilized post-surgery MRI scans ~2 months before clinically diagnosed recurrence from 50 patients with recurrent GBM. The main prediction pipeline consists of a proximity-based estimator to identify regions with high risks of recurrence (HRR), and an SVM classifier to provide voxel-wise prediction in HRR. The HRRs were estimated using the weighted sum of inverse distances to two possible origins of recurrence – SCN and tumor cavity. Subsequently, multi-parametric voxels (from T1, T1ce, FLAIR, T2, ADC) within the HRR were grouped into recurrent (warped from the clinical diagnosis) and non-recurrent subregions, and fed into the proximity estimation coupled SVM classifier - SVM_PE. The cohort was randomly divided into 40% and 60% for training and testing, respectively. The trained SVM_PE was then extrapolated to an earlier time point for earlier recurrence prediction. As an exploratory analysis, the SVM_PE predictive cluster sizes and the image intensities from the five MR sequences were compared across time to assess the progressive subclinical traces.

Results:

On 2-month pre-recurrence MRIs from 30 test cohort patients, the SVM_PE classifier achieved a recall of 0.80, a precision of 0.69, an F1-score of 0.73, and an average boundary distance of 7.49 mm. Exploratory analysis at early time points showed spatially consistent but significantly smaller subclinical clusters and significantly increased T1ce and ADC values over time.

Conclusion:

We demonstrated a novel voxel-wise early prediction method, SVM_PE, for GBM recurrence based on clinical follow-up MR scans. SVM_PE is promising in localizing subclinical traces of recurrence 2-month ahead of clinical diagnosis and may be used to guide more effective personalized early salvage therapy.

1. Introduction

Glioblastoma (GBM) is a deadly primary brain cancer that recurs in nearly all patients despite aggressive interventions, including maximal safe surgical resection, chemoradiotherapy, and tumor treating fields. Radiotherapy (RT), following surgery, offered improved survival and has long been an essential component of high-grade glioblastoma treatment [1–3]. However, RT ultimately fails to control GBM locally [4]. The unsatisfactory treatment outcome is partly because most GBMs are diagnosed in the advanced stage with tumor infiltration into the brain parenchyma that precludes a complete resection. The biological drivers of GBM resistance to treatment are not fully understood [5]. They are likely due to a combination of GBM’s intrinsic radioresistance [6], its diffusely infiltratuve nature [7–8], and tumor stem cell niche [9–10]. Overcoming tumor radioresistance requires a high radiation dose, which is contraindicated by the need for a large margin and treatment volume to encompass the diffuse tumor and stem cell niches. The predicament is further complicated because current GBM diagnosis and prognosis methods are not suited to precisely map the disease infiltration for recurrence prediction and personalized initial treatment target volume definition.

Cancergenesis analyses on tumor tissue provide a pathological snapshot of tumor cells and allow for patient profiling at the genomic and proteomic levels. However, while pathological subtyping dictates tumor aggressiveness and prognosis, it yields limited insights on the location of the recurrence for image-guided RT planning. A few radiologic efforts, including functional imaging (i.e. ¹⁸F-fluoro-ethyl-tyrosine positron emission tomography (¹⁸F-FET PET)) [11], diffusion MRI [12], CEST [13], and MR spectroscopy [14], and the new image postprocessing method [15] (i.e. radiomics analysis in peritumoral edema), have been made to predict GBM recurrence. Advanced PET and MR images taken before RT may reveal the tumor volumes inadequately covered by the RT field, but they suffer from two limitations. First, a sufficient number of cancer cells are needed to generate the imaging signal-to-noise ratio for detection. These modalities are unable to detect subclinical cancer infiltration. Second, the snapshot images at a single time may not reflect the dynamic progression of disease due to local and distal factors. Radiomics extracted features in the edema to help identify recurrence adjacent to initial tumors. However, this method cannot identify the substantial portion of the recurrence distal to the initial location. Alternatively, algorithms probing into the cell of origin and enabling early and localized detection of subclinical recurrence may offer the opportunity to treat the recurrence volume at an earlier time and to a greater radiation dose. Emerging evidence has associated the pathogenesis of GBM recurrence with the involvement of stem cell niche (SCN), providing a potential new direction for prediction [16–21]. Aiming to clarify the weak or controversial correlation between SCN and patient survival [22], a comprehensive inverse geometrical metric was developed, which revealed the statistical significance of SCN involvement in GBM recurrence and patient survival [23]. Although increasing SCN-directed studies succeeded in progression and survival risk stratification, they did not directly translate into localized recurrence prediction for salvage RT planning.

This study introduces a quantitative proximity-based support vector machine (SVM_PE) for localized recurrence prediction. We hypothesize that peritumoral infiltration and SCN migration collectively contribute to the GBM recurrence and an integrated proximity estimation to both locations provides comprehensive profiling of regional recurrence risk. Within high-risk regions (HRR) narrowed down by proximity estimation, we trained an SVM model for voxel-wise recurrence prediction on postoperative follow-up MRIs. We evaluated the performance of SVM_PE on MR scans obtained two months before recurrence and then extrapolated the model on scans acquired at earlier time points for further advanced recurrence localization.

2. Methods

2.1. Data and preprocessing

Multi-parametric MRI scans of 50 patients with recurrent GBM were retrospectively solicited from multiple databases: 9 patients from the publicly available Cancer Imaging Archive (TCIA) dataset [24] and 41 patients from two institutional databases (USC IRB #HS-19–00888 and UCLA IRB#12–001882, aged from 18–73 years (50.5 ± 13.4 years), 10 female and 31 male). Patients were selected according to the following criteria: 1) had surgical resection at initial diagnosis; 2) had clinically diagnosed recurrence; 3) received at least one follow-up MR scans within three months ahead of clinically diagnosed recurrence, as well as one diagnostic MR scan at the time of recurrence. 4) follow-up scans were taken at least one week after surgery and three months post-RT (or shown no signs of pseudo-progression). 5) MR scans consisting of 5 standard clinical MR sequences as detailed below. Additional eligible follow-up scans from 25 of the 50 subjects were included in the study for exploratory analyses. Dataset consisted of multicenter, multi-vendor MR scans with field strengths of 1.5T or 3T, and slice-thickness from 1 – 6mm. The five standard clinical sequences for each scan set included T1-weighted (T1), T1 contrast-enhanced (T1ce), T2-weighted (T2), T2 Fluid-Attenuated Inversion Recovery (FLAIR), and diffusion-weighted echo-planar imaging sequence derived apparent diffusion coefficient (ADC) map. Primary tumor volumes, defined as resection cavities in this study, and recurrence volumes were segmented by an experienced radiologist mainly on T1ce images and FLAIR, with other sequences for complimentary assessments. SCN zones, including bilateral subventricular zone extending from the wall of each lateral ventricle and subgranular zone residing at the interface between the hilus and the granular layer of each hippocampus, were manually delineated on a common T1-weighted Montreal Neurological Institute (MNI) template [23,25]. The intra-rater variability Dice score of SCN identification was 0.95, by one radiation oncology trainee spanning two weeks.

To minimize the individual variance of brain size and ventricular anatomy and thus facilitate subsequent population-based statistical analysis, we performed the following image processing. Each MRI scan was preprocessed for skull stripping, N3 bias correction, and linear alignment to the corresponding first pre-recur T1 scan [26–28]. All MR scans were resampled into the isotropic 1×1×1 mm³ resolution, and deformably registered to the MNI space using pathology adaptive deformable registration in advanced normalization tools (ANTs) [25, 29]. Specifically, the MNI template was deformably aligned to each T1 MR scan using the state-of-art Symmetric Normalization diffeomorphic registration. The noncorresponding tumor regions (i.e., cavities and recurrences) were excluded from the registration cost function. The resultant inverse transformations were then propagated onto other MR sequences and primary and recurrent tumor volumes [26, 28].

2.2. Overall pipeline

A diagram of the training scheme of the proximity-based SVM classifier, SVM_PE, is shown in Figure 1. The classifier consists of two main steps: proximity-based calculation to identify regions with a high risk of recurrence and SVM classifier training to leverage voxel-wise prediction in the high-risk areas. The SVM_PE training was conducted in the follow-up MRI scans acquired immediately before clinically diagnosed recurrence (Figure 1(ii)) when subclinical recurrence likely had started but was not discernible from clinical MRIs. The whole cohort was randomly divided into a 40% (20 subjects) training cohort and a 60% (30 subjects) testing cohort. The involved data centers were proportionally balanced for both cohorts to reduce institutional variations. As an exploratory investigation (Figure 1(ii)), the trained SVM_PE classifier was then applied on the scans acquired one time-point earlier (on average four months before recurrence) to test the feasibility of further earlier recurrence prediction.

Figure 1:

Diagram of the proximity-based SVM training pipeline (i) and the timeline of postoperative scans and the training scheme (ii). In subplot (i): (a) Generation of proximity maps to stem cell zones (P_S) and the original tumor (P_T). (b) Calculation of high-risk region-of-interest (HRR) based on (P_S) and (P_T). (c) Using non-recurrence and recurrence subregions within HRR for feature extraction on multi-parametric MRIs. (d) Training a nonlinear multi-dimensional SVM_PE using the extracted voxels. (e) The expected output of voxel-wise recurrence prediction. In subplot (ii), the SVM classifier was trained using MRI scans acquired immediately before clinically diagnosed recurrence (green icon, on average 2-month apart. Named as t-1 time point, t being the recurrence time.). The trained classifier was then tested on the further upstream follow-up scan (orange icons, on average four months before recurrence. Named as t-2 time point.) as an exploratory analysis.

2.3. Proximity estimation and High-risk region-of-interest (HRR)

We initially estimated the recurrence risks based on proximity to two hypothesized origins, SCN [23, 30–31] and original tumor (P_T) [32–33], respectively. The degree of proximity to the SCN (P_S) is estimated based on the normalized sum of inverse distances:

$$P_S(x)=\{\begin{array}{c}\frac{\mathcal{D}(x)-{\mathcal{D}}_{\mathit{min}}}{{\mathcal{D}}_{\mathit{max}}-{\mathcal{D}}_{\mathit{min}}},\mathit{if}x\notin S\\ 1,\mathit{if}x\in S\end{array}$$ (1)

where

$$\mathcal{D}(x)=\sum _{i=1}^N\frac{1}{d{\left(x,s_i\right)}^2}$$ (2)

S_i with i ∈ [1, N] denotes an observation in SCNs, with x being an interpolated voxel in the region of interest. d(x, V_i) represents the squared Euclidean distance between x and a given observation in S. ${\mathcal{D}}_{\mathit{max}}$ and ${\mathcal{D}}_{\mathit{min}}$ are the maximum and minimum inversed distances $\mathcal{D}(x)$ within S for a given subject, respectively.

Similarly, proximity estimation, P_T, to the original tumor T_i with i ∈ [1, M], is defined as:

$$P_T(x)=\{\begin{array}{c}\frac{\mathcal{D}(x)-{\mathcal{D}}_{\mathit{min}}}{{\mathcal{D}}_{\mathit{max}}-{\mathcal{D}}_{\mathit{min}}},\mathit{if}x\notin T\\ 1,\mathit{if}x\in T\end{array}$$ (3)

The high-risk region-of-interest (HRR) was estimated by the aggregated recurrence risks from the SCN and the primary tumor through a combination of P_S and P_T, as:

$$\alpha P_T+P_S\ge c$$ (4)

where α adjusts the weighting between P_T and P_S, and c thresholds the aggregated risk to create a final ROI. α and c were selected using grid search to maximize the inclusion of recurrence through a weighted mean of sensitivity and precision (maximize[Recall*2 + Precision, (ROI, Recurrence)]) on the training set. Figure 2 shows 3D renderings of calculated HRRs in 4 example subjects.

Figure 2:

3D rendering of the SCN proximity estimated HRR from step 1 for 4 example subjects. The HRRs (light blue) reflect patient-specific initial tumor location (red) and SCN proximity and effectively cover the regions with later recurrence (green), especially for patients with distal recurrences (S3 and S4.).

2.4. SVM classifier and performance evaluation

A radial basis function (RBF) kernel SVM classifier [34–35] was used to isolate recurrent subregions in previously defined HRRs further. Specifically, multi-dimensional voxels (T1, T1ce, FLAIR, T2, ADC) within HRRs were fed into SVM, with recurrent subregions (green in Figure 3) being class 1 and non-recurrent subregions excluding original tumors (blue in Figure 3) being class 2. We normalized all MR scans using the z-score method with a center of 0 and a standard deviation of 1 [36] to reduce the intensity variations due to heterogeneous acquisition protocols and scanners. To maintain a balance between the two training classes, we limited non-recurrence inputs to 30,000 voxels, randomly selected from the HRR, for each subject. An RBF kernel scale γ = 2 and a box-constraint C = 100 were selected based on the performance from a validation subset of 10 subjects from the training cohort. Due to the nature of voxel-wise prediction, the original outputs from SVM may be geometrically unrealistic (i.e., noisy and sparse), which can be alleviated by morphological postprocessing [37]. Specifically, a Gaussian kernel was applied on the binarized raw predictions from SVM, followed by distance-based clustering with minimal Euclidean distance of 1 voxel, to remove small, isolated regions whose sizes are smaller in a cutoff percentage. Kernel sizes 3×3, 5×5, 7×7, and cutoff values of 30%, 50%, 70% were investigated, and the 5×5 kernel and 50% cutoff were selected based on the performance of the training cohort.

Figure 3:

Visualization of voxel-wise prediction results from SVM_PE for two subjects from the testing cohort. Subject A demonstrates a good prediction (ABD = 1.47 mm), and subject B exemplifies a typical prediction (ABD = 8.88 mm). For both patients, the pre-recur T1ce scans overlaid with warped recurring ROI are shown on the top left, followed by the SVM_PE predictions on the bottom left, and the other four sequences (T1, T2, FLAIR, and ADC) on the right. Colors from blue to red encodes an increased pseudo recurrent risk from 0 to 1.

The SVM_PE model was first evaluated by comparing the binary classification outputs with the warped recurrence and the non-recurrence region within the HRRs, shown in Figure 1(i) as the green and light blue regions, respectively. Specifically, the performance was assessed using recall (a measurement of the classification sensitivity), precision (a measurement of the positive prediction rate), and F1 score (the harmonic mean of precision and recall). The mathematical definitions are:

$$\text{Recall}=\frac{\text{true}\text{positive}}{\text{true}\text{positive}+\text{false}\text{negative}}$$ (5)

$$\text{Precision}=\frac{\text{true}\text{positive}}{\text{true}\text{positive}+\text{false}\text{positive}}$$ (6)

$$\text{F}1=2*\frac{\text{Recall}*\text{Precision}}{\text{Recall}+\text{Precision}}$$ (7)

In addition, while the warped recurring ROI affords a higher probability of later recurrence, its specific boundaries do not necessarily dictate the demarcation of the subclinical traces. As a result, we also calculated the average boundary distance (ABD), a metric assessing the degree of spatial disagreement between the 3D surfaces, as a more generalized evaluation of the’ relevance’ of predictions [38–39]. Specifically, given a surface vertex set $S_P=\{S_{Pi},i\in \left[1,N_{S_P}\right]\}$, on the 3D surface generated from a binarized prediction, and the corresponding surface vertex set $S_R=\left\{S_{Rj},j\in \left[1,N_{S_R}\right]\right\}$, generated from the recurring ROI, ABD was obtained using:

$$ABD=\frac{1}{N_{S_P}+N_{S_R}}\left(\sum _{i=1}^{N_{S_P}}{\min }_{S_{Rj}\in S_R}d\left(S_{Pi},S_{Rj}\right)+\sum _{j=1}^{N_{S_R}}{\min }_{S_{Pi}\in S_P}d\left(S_{Rj},S_{Pi}\right)\right)$$ (8)

Where d(S_Pi, S_Rj) represents the Euclidean distance between a surface vertex S_Pi in the prediction and a surface vertex S_Rj in the corresponding recurring ROI, thus the ABD has a unit of mm.

2.5. Extrapolation to longitudinal data

To test the feasibility of early subclinical tumor tracking at a further upstream time point, we applied the trained SVM_PE model on the 25 subjects with additional MRI scans on average four months (±1.3 months) before recurrence (t-2 time point in Figure 1(ii), orange icons). Pairwise student t-tests were used to compare the SVM_PE predictive cluster sizes from two pre-recur (t-1 and t-2) time points to test detectable subclinical progressions.

As an exploratory analysis, to investigate the underline subclinical anatomical changes, normalized imaging intensities from the 5 MR sequences (T1, T1ce, T2, FLAIR, ADC) were extracted from both pre-recur time points, using the warped recurrence regions as region-of-interests. Pairwise student t-tests were then performed between the two pre-recur time-points in terms of the mean of the five extracted imaging indices.

3. Results

3.1. Two-month advanced predictions

To maximize the inclusion of recurrence in HRR, as stated in section 2.3, α = 3.5 and c = 0.9 were selected based on the training set. Metrics assessing the performance of SVM_PE for both the training and testing cohorts on 2-month pre-recur MR scans are summarized in Table. 1. Performance on the training cohort were obtained with 10-fold cross-validation. Compared to binarized recurring and non-recurring ROIs within HRR, SVM_PE resulted in averaged predictive recall of 0.81 and 80, averaged precision of 0.68 and 0.69, and averaged F1 score of 0.73 and 0.73 for the training and the testing cohort, respectively. Group-wise t-tests revealed no significant performance differences between the two cohorts. Similar performance between the two cohorts was also confirmed by the distance-based metric ABD, indicating good generalization of SVM_PE onto the test set. Figure 3 shows SVM_PE predictions for two example subjects and the corresponding 5 MR scans in the same view. For both subjects, the demarcation of T1 contrast-enhanced regions greatly underestimated the regions with later recurrence (red, warped, and overlaid on T1ce images). At the same time, intensity enhancements in T2, FLAIR, or ADC maps were less specific in revealing subclinical traces. In contrast, the SVM_PE predictions corresponded well with later clinical diagnosed recurrence in both cases.

Table 1:

Performance evaluation metrics of SVM_PE on the 2-month pre-recur dataset. For both the training and testing datasets, four metrics (Recall, Precision, F1 score, and ABD) are presented, along with their corresponding group-wise comparison p-values. No statistically significant (p < 0.05) differences were detected on the four metrics between the discovery and testing sets.

	Training*		Testing		P-values
	Mean	SD	Mean	SD
Recall	0.81	0.13	0.80	0.10	0.591
Precision	0.68	0.14	0.69	0.14	0.684
F1	0.73	0.12	0.73	0.10	0.978
ABD (mm)	8.14	6.33	7.49	6.19	0.722

^*For the training dataset, performance scores are reported after 10-fold cross-validation.

3.2. Longitudinal subclinical changes

Results of longitudinal predictions and intensity comparisons across time are shown in Figures 4 and 5. As we can see from Figure 4, compared to the prediction from the immediate pre-recur (2 months prior) time point, the predicted clusters from 4-month apart were located in similar locations but generally smaller in size. This pattern can be evidently visualized in Figure 5(a), where the majority of the subjects presented an enlargement in subclinical regions as approaching the recurrence. The pairwise t-test revealed a significant size difference between the two time points, with p = 0.004. Besides, comparisons of image indices on T1ce and ADC maps between these two time points revealed significant intensity enhancements over time, while comparisons on T1, T2, and FLAIR signals failed to reach group-wise significance.

Figure 4:

Visualization of SVM_PE predictions on the 4-month (two columns on the left) and 2-month (two columns on the right) pre-recur time points for 4 example subjects. SVM predictions are shown next to the corresponding T1ce scan from the same time point for all patients. Warped recurrences are shown in red and overlaid on the T1ce scans. Colors from blue to red encodes an increased pseudo recurrent risk from 0 to 1.

Figure 5:

Results of exploratory comparisons between two pre-recur time points on SVM_PE predictive cluster size, as well as 5 MR image intensities within recurrence ROIs. Statistically significant (p < 0.05) differences between the two time points were detected in cluster sizes, T1ce intensities, and ADC values.

4. Discussion

In the study, we showed the feasibility of predicting the subclinical recurrence locations based on post-RT longitudinal MR images and biology-driven geometrical features. The novel therapeutic implication and voxel-wise recurrence prediction method are distinctly different from existing efforts of recurrence prediction.

Existing research for recurrence prediction is intended for modification of the initial treatment volume. For this reason, the studies are focused on the pretreatment images. Peritumoral infiltration of tumor cells has been considered one of the primary sources of GBM recurrence, with edema areas having the highest probability of infiltration [32]. Based on this assumption, Rathore et al. performed a radiomics analysis on peritumoral edema using pre-operative traditional and perfusion MRI and achieved decent prediction in local recurrence [15]. Nonetheless, although pre-operative local infiltration may contribute to later recurrence, especially the ones that recurred within a few months following treatment, it is not the only resource of recurrences, as a substantial portion of the recurrence spread into regions distant (> 2cm) from the initial location [30]. Without the peritumoral assumption, Harat et al. prospectively investigated the possibility of using 18F-fluoro-ethyl-tyrosine (18F-FET) PET to predict the recurrence location [11], as the MR-defined target volume insufficiently covered the PET avid volume. Although the FET avid volumes showed improved overlap with the recurrent volumes, they suffer from two limitations. First, the number of cells (i.e., 10E6 cells) required to produce sufficient PET signals prevents it from detecting microscopic infiltration [40]. Second, the pretreatment snapshot cannot capture the morphological and pathological alterations during and after initial treatments [41]. Other advanced imaging methods for recurrence prediction, including diffusion MRI [12], CEST [13], and MRS [14], suffer from similar limitations. Unlike these studies aiming to modify the initial treatment volume, we designed our study to improve the recurrent treatment by fully utilizing the longitudinal imaging information for patients under frequent imaging surveillance following their initial treatment. Early GBM recurrence prediction provides a second avenue for the personalized therapeutic design, which may be implemented at an earlier time point to maximally delay recurrent disease progression with safe radiation dose escalation to a smaller target volume.

Regarding the prediction method, beyond what can be captured in images of the tumor and its periphery, recent cancerogenesis analyses have suggested the role of SCN in the pathogenesis of GBM recurrence, providing a new direction for localizing subclinical trace of recurrence [16]. For example, recurrence has been demonstrated to have a higher tendency to spread closer to SCN, regardless of the initial tumor location [3, 23, 42]. However, the radiologic definition of SCN is inconsistent among studies [3,18,42]. The proximity to SCN has been ineffectively characterized by the semi-quantitative measurements adopted by most of the SCN derived studies. For the same reason, the role of SCN in GBM formation and progression has not been translated into localized voxel-level recurrence prediction.

Our localized GBM recurrence prediction model is based on the interplaying contribution of peritumoral infiltration and SCN migration, thus providing comprehensive profiling of regional recurrence risk. On clinical follow-up MRI, the proximity to the SCN and primary tumor was quantitatively characterized using inverse distance mapping. The GBM localization was consecutively narrowed down by integrated proximity estimation and voxel-wise SVM classification. On 2-month pre-recur scans, our model successfully delineates subclinical pathologies that are consistent in the location with later radiologically confirmed recurrences, albeit not clearly defined from either single MR sequence (Figure 3). The performance comparison between the training and testing sets suggests the good generalizability of the SVM_PE model (Table 1). Since our SVM model is based on standard clinical MR sequences, the proposed method will be easily implemented as an add-on to current clinical practice without additional sophisticated imaging scans.

In extrapolation analysis to an earlier time point (4 months before recurrence), the trained SVM_PE model predicted significantly smaller cluster size in similar locations, indicating a potential trend of subclinical progression and the possibility of using SVM_PE for more advanced subclinical trace localization. The existence of progressive subclinical changes is further evidenced by the exploratory analysis of MR signals between two pre-recur time points (Figure 5). In particular, as the current gold standard of clinical tumor recurrence diagnosis, T1ce signals presented a small but significant increase from 4 to 2 months before recurrence. On top of it, the considerable rise of ADC values, which are inversely proportional to tumor cellularity, may indicate a decreased microstructural neurite density [43]. These are in line with the findings from Jin et al., where neurite density loss, represented by changes in fractional anisotropy and neurite density indices, was also observed at two months before recurrence on 30 GBM patients [44]. Our novel method captures the SCN and MR information more comprehensively for early recurrent location prediction.

We evaluated our model using both volume and average boundary distance agreement between the prediction and the actual recurrence. The performance indicates the ability to locate future recurrence, but the factor of tumor progression between the time points needs to be considered. A lower recall is expected, particularly for further upstream analysis where the predicted tumor volume is smaller than recurrence at a later time point. Besides performance evaluation, the distance agreement can help define the target volume in radiotherapy planning.

Our study has several notable limitations. First, the use of generic-lesion-based registration and the MNI template for brain alignment. Although being the current state-of-the-art method for pathological brain registration, lesion-based ANTs cannot fully adapt to specific challenges in GBM brain alignment, i.e., mapping excessive edema that constantly changes across the follow-up period. On top of it, the MNI brain template based on healthy subjects inadequately captures anatomy characteristics in GBM brains, such as degenerated white matter and enlarged lateral ventricles. These differences may introduce further errors in the manual identification of SCN, especially the areas in ventricular vicinity. In the future, an advanced pathological-adaptive registration algorithm and a brain template specific to the GBM population are needed to facilitate more accurate intra- and inter-patient image alignment [41]. Second, the generation and testing of the SVM_PE were based on a limited dataset without screening on confounding factors, such as genomic subtypes initial tumor sizes, and treatment regimens. The moderate standard deviation on prediction performance (Table. 1) may suggest a non-negligible population heterogeneity, possibly due to the underlying divergent recurrent patterns and speeds. Whether the performance differences dictate pathological or therapeutic subtypes and whether the current model can be further refined according to these subtypes need to be explored in a larger dataset. Third, the MR datasets were obtained through multi-scanners on multi-sites. While normalized using a standard protocol [36], the remarkable variations in tissue intensity due to acquisition protocol and scanner differences can not be disregarded. In the future, we will prospectively acquire quantitative MR images to minimize the impact of inter-scan variations on SVM predictions [45]. Fourth, the SVM_PE model was trained and tested on scans sparing the 3-month time window after RT to avoid the complications of pseudo-progression. However, pseudo-progression is commonly seen in 20–30% of patients following RT and remains a critical factor in interfering with clinical decisions [43, 46]. Therefore, although the current model yield localized early recurrence prediction two months before clinical diagnosis, an advanced prediction model capable of identifying the distinctive trace of recurrence over concurrent treatment-related pseudo-progression are needed in the future, to achieve clinically robust recurrence prediction and monitoring in the entire clinical course after treatment [47].

5. Conclusion

In the presented study, we introduced a novel voxel-wise early prediction method for brain GBM recurrence based on clinical follow-up MR scans using an SCN proximity estimation coupled SVM framework- SVM_PE. The result shows promise in localizing subclinical traces of recurrence 2-month ahead of clinical diagnosis. The proposed method could be easily implemented as an add-on to current clinical practice without additional imaging scans. The derived voxel-wise prediction may serve as a target volume for advanced dosimetric planning that maximally includes invisible infiltrative cancer cells in the recurrent treatment at an earlier time to delay cancer progression.