Prediction of post-operative deficits using an improved diffusion weighted imaging maximum a posteriori probability analysis in pediatric epilepsy surgery

Abstract

OBJECTIVE

This study is aimed to improve the clinical utility of diffusion weighted imaging maximum a posteriori probability (DWI-MAP) analysis previously reported to be useful for predicting postoperative deficits in motor, language and visual field in pediatric epilepsy surgery. We determined the additive value of a new clustering mapping method in which average direct-flip distance (ADFD) re-classifies the outliers of original DWI-MAP streamlines by referring to their minimum distances to the exemplar streamlines (i.e., medoids).

METHODS

We studied 40 children with drug-resistant focal epilepsy (age: 8.7±4.8 years) who underwent resection of the presumed epileptogenic zone and had five categories of postoperative deficits (i.e., hemiparesis involving face/hand/leg, dysphasia requiring speech therapy, and/or visual field cut). In pre- and postoperative images of the resected hemisphere, DWI-MAP identified a total of nine streamline pathways, C₁: face motor, C₂: hand motor, C₃: leg motor and C₄: Broca’s area-Wernicke’s area, C₅: premotor area-Broca’s area, C₆: premotor area-Wernicke’s area, C₇: parietal area-Wernicke’s area, C₈: premotor area-parietal area, and C₉: occipital lobe-lateral geniculate nucleus. For each streamline of the identified pathway, the minimal ADFD to nine exemplars corrected the pathway membership. Binary logistic regression analysis was employed to determine how accurately two fractional predictors: Δ_1–9 (postoperative volume change of C_1–9) and γ_1–9 (pre-operative volume of C_1–9 surgically resected) predicted postoperative motor, language and visual deficits.

RESULTS

Addition of ADFD to DWI-MAP improved the sensitivity/specificity of regression models for predicting postoperative motor, language and visual deficits, Δ_1–3: 28% (from 0.62 to 0.79), Δ_4–8: 13% (from 0.69 to 0.79), Δ₉: 13% (from 0.77 to 0.87), γ_1–3: 7% (from 0.81 to 0.87), γ_4–8: 1% (from 0.86 to 0.87) and γ₉: 24% (from 0.75 to 0.93). Preservation of the eloquent pathways defined by preoperative DWI-MAP with ADFD (up to 97% of C_1–4,9) prevented postoperative motor, language and visual deficits with sensitivity/specificity ranging from 88 to 100%.

CONCLUSIONS

The present study suggests that postoperative functional outcome substantially differs according to the extent of resected white matter encompassing eloquent cortex determined by preoperative DWI-MAP analysis, underlying that the preservation of preoperative DWI-MAP pathways may be crucial to prevent postoperative deficits. The improved DWI-MAP analysis may provide a complementary non-invasive tool capable of guiding the surgical margin to minimize the risk of postoperative deficits for children.

Introduction

Clinical management of medically intractable epilepsy involves surgical resection of a patient’s epileptogenic zone, with the goal of preventing future seizure onset and minimizing postoperative functional deficits.¹⁸ Successful minimization of postoperative deficits depends on the accurate preoperative mapping of patient’s “eloquent” cortical areas associated with primary motor, language and visual functions.^9,28,29 Unfortunately, this mapping is often challenging in children who may not satisfactorily cooperate during electrical stimulation mapping (ESM) or functional MRI (fMRI) studies.^19,36 Furthermore, it remains unclear what volume fraction of eloquent brain area including white matter may be resected without causing acute and clinically-evident postoperative deficits, or whether there exists a specific fractional threshold of resected eloquent cortex associated with such postoperative deficits.

In the last two decades, diffusion weighted imaging (DWI) tractography has been recognized as a powerful technique to visualize white matter pathways at the whole-brain level, even with minimal patient cooperation.^4,8,10,11,30 Especially, in our recent studies of children with focal epilpesy,^20–22 a maximum a posteriori probability (MAP) classification method was able to detect eloquent pathways supporting primary motor and language functions(with high accuracy for streamlines terminating in proximity to the localization of ESM, which is the current gold-standard to identify eloquent tissues in children who undergo epilepsy surgery: 77% for face, 76% for fingers, 86% for leg, 60% for language (contact), and 89% for face, 89% for fingers, 88% for leg, 90% for language (surrounded within 2 cm). This approach was originally designed in the Bayesian classification framework to classify a given streamline into one of multiple target classes by calculating MAP estimates of individual fiber streamlines from fMRI-derived pathway probability maps.^20,22 A major advantage of this approach is its simultaneous localization of eloquent pathways at both the cortical and subcortical levels without requiring additional mapping modalities such as fMRI and ESM.

The present study further investigates whether an additional clustering procedure using the pathway exemplars (i.e., the most representative pathway streamlines obtained from healthy controls) may reduce false positive fibers of original DWI-MAP classification leading to the improvement of original DWI-MAP approach to predict postoperative deficits. To this end, we utilize the QuickBundles algorithm,¹³ in which the class memberships of original DWI-MAP pathways are corrected to minimize the degree of fiber dissimilarity. Our specific hypothesis is that the DWI-MAP pathways corrected by the QuickBundles algorithm will yield a more accurate prediction of postoperative deficits compared to the original DWI-MAP pathways. We systematically investigate the relationship of the postoperative deficits with 1) the postoperative DWI-MAP pathway volume changes, and 2) the resected fraction of preoperative DWI-MAP pathway volumes.

Methods

Subjects

We recruited 96 children with intractable focal epilepsy (age: 7.1±5.2 years, 54 boys) who underwent cortical resection of the presumed epileptogenic zone between 2009 and 2017 via two-stage epilepsy surgery protocol of Children’s Hospital of Michigan or Harper University Hospital, Detroit, MI, USA. Among those children, Thereby, the present study included all 40 children who had underwent both preoperative and postoperative MRI evaluations (time interval: 11.2±7.3 months, age: 8.7±4.8 years, 18 boys) and satisfied the following enrollment criteria. The inclusion criteria were: 1) drug resistant focal epilepsy requiring two-stage epilepsy surgery with chronic subdural ESM mapping, and 2) no preoperative motor or language impairment. The exclusion criteria were 1) history of prematurity or perinatal hypoxic-ischemic event (n=5), 2) hemiplegia on preoperative examination by pediatric neurologists (n=5), 3) dysmorphic features suggestive of a clinical syndrome (n=4), and 4) MRI-visible massive brain malformations that distort the central or lateral sulcus (n=6, because essential language function is likely to be present in the opposite hemisphere), 4) No postoperative MRI data scanned (n=40). For each participant, the resection margin was determined as previously described.³¹ The presence of acute and newly developed postoperative functional deficits were determined clinically by a pediatric neurologist (A.L), as well as physical, occupational, and speech therapists within two weeks from resective surgery. All clinical team members were blinded to the proposed DWI data analysis. Deficits were categorized as (i) face motor deficit, (ii) hand motor deficit, (iii) leg motor deficit, (iv) dysphasia, and/or (v) visual field cut. Postoperative seizure outcome was evaluated using ILAE classification every 6 months either at clinical visit or using phone interview (i.e., ILAE class I/II/III/IV/V: 24/2/4/8/2 reported at least 1year after surgery).

Table 1 presents the patient profiles. This study was performed in accordance with policies of the Wayne State University Institutional Review Board, and a written consent was obtained from the patient’s guardian. All data were obtained as a part of routine clinical management of patients. Thus, no prospective intervention was performed by the results of the presented DWI-MAP method to ensure no additional, direct risk of harm from this study to the patient beyond the risk of the clinical surgery procedures. That is, no recommendation was made by this study to guide the surgery plan since all DWI-MAP analyses were retrospectively performed after surgery.

Table 1.

Clinical variables of the recruited study subjects.

No	Age	Gender	Side	Number of AED	Daily seizure	Resection lobe	Postoperative deficit
1	12	G	Lt	2	No	T	No
2	5	G	Rt	3	Yes	Fr,T,P,O	Fa,H,L,La,V
3	17	B	Lt	2	No	T	No
4	14	B	Lt	2	No	P	No
5	9	G	Lt	2	Yes	T	La,V
6	12	G	Lt	2	Yes	Fr,P	Fa,H,L,La
7	7	B	Rt	3	Yes	T,P,O	H,V
8	17	B	Rt	2	No	O	V
9	12	B	Lt	2	Yes	P	No
10	11	G	Lt	2	Yes	T	No
11	6	G	Rt	1	No	Fr	No
12	4	B	Rt	1	Yes	T	No
13	11	G	Rt	2	Yes	Fr,T,O	Fa,H,L,V
14	14	G	Rt	2	No	P	V
15	10	B	Rt	3	Yes	P	No
16	16	B	Lt	2	No	Fr	No
17	8	B	Rt	2	Yes	Fr	No
18	5	G	Lt	3	Yes	Fr,P,T	Fa,La,V
19	17	B	Lt	3	No	T	No
20	13	G	Rt	3	Yes	T	No
21	5	G	Lt	1	Yes	T	No
22	11	G	Rt	2	Yes	Fr,T	No
23	14	G	Lt	3	No	P	No
24	11	G	Lt	1	No	T,P,O	H,La,V
25	2	B	Lt	3	Yes	T	No
26	3	G	Lt	2	Yes	T,P,O	V
27	11	G	Lt	2	Yes	T	No
28	7	G	Lt	1	No	Fr	No
29	12	G	Rt	1	No	T	No
30	3	G	Lt	3	No	T	No
31	2	B	Rt	2	Yes	Fr,T,P,O	Fa,H,L,V
32	3	B	Rt	3	Yes	Fr,T,P,O	Fa,H,L,V
33	13	B	Lt	2	No	T	V
34	2	G	Rt	3	No	Fr,T,P,O	Fa,H,L,La,V
35	2	G	Rt	1	Yes	Fr,T,P,O	No
36	10	G	Rt	2	No	T	No
37	7	B	Rt	3	Yes	Fr	No
38	8	B	Rt	1	Yes	Fr,T,P,O	No
39	1	B	Lt	3	Yes	Fr	No
40	2	B	Rt	3	No	Fr,P	No

Age in years. G/B: girl/boy, Lt/Rt: left/right, Fr/T/P/O: frontal/temporal/parietal/occipital, Fa/H/L/La/V: face/hand/leg/language/vision.

Data analysis

All DWI scans were performed on a 3T MRI scanner using 55 isotropic gradient directions with b = 1000 s/mm². Detailed DWI parameters and whole brain tractography procedure are available in our previous study²³. For each patient, we first sorted out whole brain tract streamlines using a DWI-MAP classifier^20,22 to identify 9 functionally important white matter pathways in each hemisphere: “C₁: face motor area-internal capsule,” “C₂: hand motor area-internal capsule,” “C₃: leg motor area-internal capsule,” “C₄: Broca’s area (BA 44/45)-Wernicke’s area (BA 22),” “C₅: premotor area (BA 6)-Broca’s area,” “C₆: premotor area-Wernicke’s area,” “C₇: parietal area (BA 39)-Wernicke’s area,”, “C₈: premotor area-parietal area” and “C₉: occipital lobe-lateral geniculate nucleus”. Briefly, this DWI-MAP classifier utilized stereotaxic probability maps of C_i obtained from age-handedness matched healthy children^20,22. To classify a given streamline into C_i, a posteriori probability to belong to a class C_i was calculated by averaging the probability values of C_i over the entire trajectory under equal class prior assumption of Bayesian inference. The argument of i having the maximum a posteriori probability determined final class membership where ad-hoc equal class prior assumption causes inevitable false positive errors^15,16, especially in classifying streamlines of which trajectories are partially overlapped in multiple stereotaxic probability maps.

To reduce false positive classification in original DWI-MAP analysis, the present study adopted additional streamline clustering procedure based on the average direct-flip distance (ADFD, mean distance between equally sampled points of bidirectional fibers to quantify their similarity in geometrical location) using the QuickBundles algorithm¹³ in which original DWI-MAP streamlines, C_i, were resampled at equidistant points and then reclassified by referring their minimum ADFDs to the class exemplars. Supplemental Fig. 1 shows the most representative streamline patterns of C_i, called “exemplars”, which are mathematically centroid streamlines of C_i in healthy controls: C_i obtained from the original DWI-MAP classifier^20,21. For each streamline of the identified class C_i, ADFDs to the exemplars were evaluated in standard template space where the class membership of an exemplar having minimum ADFD less than an optimized threshold, β, determines the new class membership. This DWI-MAP procedure corrected by ADFD threshold β was used to detect nine separate white matter pathways, C_i=1–9, for each resected hemisphere in both pre- and postoperative DWI data.

Finally, two streamline volume percentages of the corrected class, C_i, were evaluated to predict the presence of postoperative deficit in subsequent statistical analysis:

$${\Delta }_i=100\times \frac{\left(volumeofpreoperative{C}_i-volumeofpostoperative{C}_i\right)}{volumeofpreoperative{C}_i}$$

$${\gamma }_i=100\times \frac{(volumeofpreoperative{C}_i\cap volumeofresectedtissue\text{) }}{volumeofpreoperative{C}_i}$$

where ∩ indicates the intersection. To determine the fractional volume of resected tissue, the resected area as seen on postoperative DWI was first manually segmented and then registered to preoperative DWI using the symmetric diffeomorphic image normalization algorithm provided through the Advanced Normalization Tools (ANTs) package.³

Statistical analysis

A binary logistic regression model was combined with receiver operating characteristic (ROC) analysis in order to examine if both Δ and γ outperform the following clinical predictors of postoperative deficit (Table 1): 1) the number of antiepileptic drugs (AED) taken before surgery, which might require larger cortical resection for subsequent seizure control,^25,31 2) seizure frequency, dichotomized as 1 if seizures occurred at least daily and as 0 if seizures were less frequent, 3) surgery side, dichotomized as 1 for right hemisphere and 0 for left hemisphere. In each predictor model, to detect the presence of postoperative deficit, the output was categorized as a group (presence of postoperative deficit = 1, no postoperative deficit = 0), and assumed as a linear function of the predictor with a subject-specific random intercept. The regression coefficient, b, of the individual predictor providing the highest pseudo correlation between the model and actual output (i.e., Nagelkerke R2) was found using a generalized linear model with binomial logit distribution. For each predictor, the present study utilized a generalized pattern search algorithm² to optimize the ADFD threshold of β, providing the most significant p-value of b. Sensitivity and specificity of the model to predict the occurrence of postoperative deficit was evaluated using ROC curve analysis.²⁷ The Δ and γ values maximizing the sensitivity and specificity in this analysis (Δ_cutoff and γ_cutoff) were taken as the optimal threshold of the predictor for a deficit-free postoperative outcome across subjects. Finally, Fisher’s exact test¹ was applied at the obtained threshold to assess statistical significance in detecting postoperative deficits.

Results

Across all subjects, the following deficits were noted postoperatively: face motor deficit (n = 7, 17.5%), hand motor deficit (n = 7, 17.5%), leg motor deficit (n = 6, 15%), dysphasia (n = 6, 15%) and visual field deficit (n = 13, 33%). Resection volume differed by the side of surgery (mean volume: 120.2/40.6 × 1000 mm³ for left/right surgery; p = .008). Total resection volume correlated with age at surgery (Spearman’s rho = −0.34, p = .03) and also the presence of postoperative deficit (Spearman’s rho = 0.56/0.76/0.91/0.38/0.70, p = .000/.000/.000/.015/.000 for face, hand, leg, language and visual field, respectively).

Table 2 shows the significance of the proposed predictors in detecting the presence of postoperative deficits using a binary logistic regression model. Compared to both clinical predictors and original DWI-MAP predictors, the improved DWI-MAP predictors were more significantly associated with the postoperative deficit, p = .009/.010/.002/.005-.020/.000 (Δ₁/Δ₂/Δ₃/Δ_4–8/Δ₉), demonstrating their notable linear relationship with the likelihood of postoperative deficits in primary motor and language functions. It should be noted that the present study optimized the ADFD threshold of β for each C_i such that the resulting Δ_i=1–9 maximize their predictability of postoperative deficit, as presented in Supplemental Fig. 2.

Table 2.

Statistics of binary logistic regression analyses associated with individual predictors to estimate the probability of post-operative deficits.

Predictor	Post-operative deficit
	Face weakness		Hand weakness		Leg weakness		Language		Visual field
	R2	p-value	R2	p-value	R2	p-value	R2	p-value	R2	p-value
Number of AED	0.12	0.107	0.04	0.334	0.08	0.215	0.02	0.506	0.03	0.345
Seizure frequency	0.12	0.127	0.05	0.274	0.09	0.194	0.01	0.624	0.00	0.720
Surgery side	0.07	0.226	0.10	0.130	0.14	0.108	0.04	0.384	0.04	0.315
Δ1,DWI-MAP	0.12	0.105	0.14	0.069	0.28	0.022*	0.04	0.356	0.09	0.110
Δ2,DWI-MAP	0.09	0.159	0.03	0.438	0.13	0.100	0.12	0.113	0.05	0.260
Δ3,DWI-MAP	0.23	0.030*	0.19	0.043*	0.27	0.021	0.08	0.185	0.07	0.189
Δ4,DWI-MAP	0.18	0.055	0.16	0.066	0.14	0.100	0.64	0.007	0.11	0.094
Δ5,DWI-MAP	0.25	0.024*	0.21	0.033*	0.23	0.026*	0.41	0.015*	0.13	0.077
Δ6,DWI-MAP	0.02	0.529	0.01	0.636	0.02	0.553	0.07	0.413	0.09	0.153
Δ7,DWI-MAP	0.21	0.032*	0.35	0.019*	0.24	0.027*	0.51	0.026*	0.30	0.024*
Δ8,DWI-MAP	0.14	0.609	0.02	0.776	0.01	0.796	0.16	0.715	0.08	0.543
Δ9,DWI-MAP	0.28	0.012*	0.53	0.002*	0.38	0.011*	0.12	0.082	0.60	0.000*
Δ1, DWI-MAP+ ADFD	0.43	0.009*	0.41	0.007*	0.56	0.012*	0.18	0.053	0.30	0.009*
Δ2,DWI-MAP+ADFD	0.42	0.010*	0.39	0.010*	0.58	0.022*	0.21	0.035*	0.10	0.105
Δ3,DWI-MAP+ADFD	0.52	0.002**	0.55	0.002**	0.60	0.002**	0.36	0.007	0.34	0.006*
Δ4,DWI-MAP+ADFD	0.04	0.317	0.07	0.185	0.05	0.294	0.48	0.005	0.02	0.486
Δ5,DWI-MAP+ADFD	0.37	0.005**	0.12	0.083	0.26	0.018*	0.36	0.007*	0.13	0.060
Δ6,DWI-MAP+ADFD	0.13	0.079	0.18	0.038*	0.23	0.029	0.26	0.022*	0.01	0.695
Δ7,DWI-MAP+ADFD	0.20	0.038*	0.32	0.012*	0.17	0.061	0.48	0.009*	0.23	0.020*
Δ8,DWI-MAP+ADFD	0.21	0.041*	0.37	0.007	0.32	0.017*	0.30	0.019*	0.11	0.123
Δ9,DWI-MAP+ADFD	0.29	0.012*	0.54	0.002*	0.39	0.011*	0.14	0.081	0.63	0.000*

R²: Nagelkerke R², * and ** indicates p-value < 0.05 and 0.005, respectively.

Figure 1 presents representative examples of preoperative DWI-MAP pathways C_1–9, which were obtained from the original and improved analysis of a 10-year-old child (patient no.15). As marked by white arrows, the original analysis yielded outlier tracts that deviated from cortical centroids of C₂ (hand motor), C₅ (premotor-Broca), C₇ (parietal-Wernicke), and C₉ (occipital lobe-lateral geniculate nucleus) which could be appropriately deleted from their pathways by using the improved analysis with an optimized ADFD threshold (actual values are reported in the legend). Interestingly, the improved analysis could identify more dense tracts in the C₆ (premotor-Wernicke) pathway, an effect that had not been fully detected in the original analysis. This suggests that the improved analysis may then provide more overlap between DWI-detected cortical terminals of eloquent white matter pathways and cortical localization of function by ESM. Figure 2 presents illustrative examples where compared with original DWI-MAP, with ADFD-correction with β* provides better spatial-match between cortical terminals of C₁/C₄ and ESM in somatosensory mapping of face/ expressive aphasia with naming impairment.

Figure 1.

Representative examples of C_1–9 pathways obtained from original DWI-MAP (left column) and improved DWI-MAP (right column, the optimized ADFD threshold β* = 16 mm, 9 mm, 8 mm,14 mm,12 mm, 16 mm,13 mm, 13 mm, 13 mm for C_1–9, respectively). White arrows indicate false positive fibers (i.e., outliers) in the original DWI-MAP which were appropriately corrected by the improved DWI-MAP.

Figure 2.

Spatial overlap between cortical terminals of DWI-MAP pathways and cortical mapping by ESM. Terminals from preoperative C₁: face motor pathway (patient no.18 with postoperative face motor weakness) and C₄: Broca’s area-Wernicke’s area pathway (patient no.18 with postoperative language deficit) were compared with their ground truth locations determined by ESM.

Figure 3 demonstrates the efficacy of ADFD correction on DWI-MAP analysis by highlighting cases where major streamlines were appropriately reclassified. Application of ADFD resulted in increased Δ for C₂, predicting higher risk of postoperative hand motor deficit in Patient no.13 (who indeed had postoperative hand weakness), resulted in decreased Δ for C₄, predicting lower risk of postoperative language deficit in patient no. 9 (who indeed had no postoperative dysphasia), and resulted in increased Δ for C₉, predicting higher risk of postoperative visual field deficit in Patient no. 8 (who indeed had postoperative visual field cut). This provides evidence that the improved analysis helps address the outlier problem presented in Fig. 1, and that it does so in a way that can be clinically useful in predicting deficit after surgery.

Figure 3.

Representative examples of C₂: hand motor pathway (patient no.13 with postoperative hand motor weakness), C₄: Broca’s area-Wernicke’s area pathway (patient no.9 with no postoperative language deficit) and C₉: occipital lobe-lateral geniculate nucleus (patient no. 8 with postoperative visual field deficit), obtained from a set of preoperative (left, blue fibers) and postoperative (right, red fibers) DWI analyses using the original DWI-MAP (top) and the improved DWI-MAP (bottom, DWI-MAP+ADFD). For both original and improved DWI-MAP pathways, Δ indicates postoperative fiber volume change normalized by preoperative fiber volume.

Comparison of prediction ability between the original and improved analysis was evaluated by a separate ROC analysis presented in Table 3. The improved analysis outperforms the original analysis in predicting postoperative deficits for primary motor, langue and visual functions. Average area under curve (AUC) for primary motor pathway (Δ_1–3)/language pathway (Δ_4–8)/visual pathway (Δ₉) improved from 0.61±0.07/0.71±0.31/0.824 to 0.85±0.08/0.83±0.07/0.910 when adding ADFD-correction to the analysis. This corresponded to equal improvements in sensitivity and specificity, with primary motor pathway (Δ_1–3) average sensitivity/specificity increasing from 0.62±0.05 to 0.79±0.07, language pathway (Δ_4–8) average sensitivity/specificity increasing from 0.69±0.24 to 0.78±0.05, visual pathway (Δ₉) sensitivity/specificity increasing from 0.771 to 0.866. Fisher’s exact test found that the improved analysis could achieve a more significant association between the predicted deficit and true deficit (average p = .04 and .28 for the improved and original analysis, respectively).

Table 3.

Accuracy of DWI-MAP and DWI-MAP with ADFD (DWI-MAP+ADFD) to predict post-operative deficit in face weakness (Δ₁), hand weakness (Δ₂), leg weakness (Δ₃), language (Δ_4–8) and visual (Δ₉).

DWI-MAP	Predictor
	Δ1	Δ2	Δ3	Δ4	Δ5	Δ6	Δ7	Δ8	Δ9
Area under curve	0.654	0.532	0.657	0.936	0.940	0.656	0.835	0.191	0.824
Asymptotic p-value	0.104	0.392	0.114	0.000	0.000	0.116	0.001	1.000	0.001
Δcut-off	34.114	28.7	18.5	36.1	12.3	56.0	14.1	54.5	21.3
Sensitivity	0.650	0.557	0.650	0.873	0.876	0.638	0.767	0.296	0.771
Specificity	0.651	0.558	0.650	0.872	0.875	0.639	0.768	0.296	0.772
Fisher’s exact test p-value	0.407	0.709	0.654	0.000	0.000	0.668	0.055	0.021	0.023
DWI-MAP+ADFD	Δ1	Δ2	Δ3	Δ4	Δ5	Δ6	Δ7	Δ8	Δ9
Area under curve	0.820	0.787	0.939	0.913	0.872	0.742	0.850	0.779	0.910
Asymptotic p-value	0.001	0.002	0.000	0.000	0.000	0.024	0.000	0.009	0.000
Δcut-off	32.690	28.9	30.9	43.4	18.4	20.4	34.5	36.1	24.5
Sensitivity	0.767	0.742	0.870	0.846	0.800	0.720	0.783	0.737	0.866
Specificity	0.767	0.743	0.871	0.846	0.800	0.720	0.784	0.737	0.866
Fisher’s exact test p-value	0.027	0.036	0.002	0.004	0.039	0.159	0.010	0.075	0.000

^*Not available (N.A) cut-off threshold since five predictors were combined.

To investigate the potential efficacy of this DWI-MAP analysis in predicting postoperative deficit using preoperative DWI data alone, the volumetric fraction of preoperative DWI-MAP pathways resected by surgery (i.e., γ_1–9) was tested as a new predictor in the binary logistic regression model combined with ROC analysis (Table 4). It should be noted that each γ_1–9 outperformed the corresponding Δ_1–9 to predict the presence of postoperative deficits in both the original and improved analysis, supporting the clinical utility of the preoperative DWI-MAP analysis in predicting postoperative deficits given the fraction of a pathway that was (or will be) resected. Note that the preservation of preoperative pathways C_1–4,9 at the reported cut-off values (γ_cut-off) yields a predicted 88%, 85%, 93%, 99% and 93% of patients without deficit in face motor, hand motor, leg motor, language and visual function, respectively, at Fisher’s exact test with p < .007. Fisher’s exact test found that up to 3% resection of DWI-MAP-defined pathways (i.e. γ_cut-off) would provide the least likelihood of postoperative deficit in face motor (γ₁: TP = 6/7, FN=29/33), hand motor (γ₂: TP = 6/8, FN=32/32), leg motor (γ₃:TP = 6/6, FN=32/34), language (γ₄: TP = 6/6, FN=33/34) and visual field (γ₉: TP = 13/13, FN=25/27).

Table 4.

Accuracy of DWI-MAP and DWI-MAP with ADFD (DWI-MAP+ADFD) to predict post-operative deficit in face weakness (γ₁), hand weakness (γ₂), leg weakness (γ₃), language (γ_4–8) and visual (γ₉).

DWI-MAP	Predictor
	γ1	γ2	γ3	γ4	γ5	γ6	γ7	γ8	γ9
Area under curve	0.758	0.846	0.965	0.996	0.867	0.910	0.944	0.747	0.785
Asymptotic p-value	0.011	0.000	0.000	0.000	0.000	0.000	0.000	0.022	0.012
γcut-off	0.2	0.3	0.4	3.0	0.1	0.2	0.9	0.5	0.2
Sensitivity	0.734	0.790	0.908	0.985	0.829	0.869	0.881	0.713	0.746
Specificity	0.734	0.789	0.909	0.985	0.830	0.869	0.880	0.714	0.747
Fisher’s exact test p-value	0.001	0.009	0.000	0.000	0.0004	0.001	0.002	0.005	0.001
DWI-MAP+ADFD	γ1	γ2	γ3	γ4	γ5	γ6	γ7	γ8	γ9
Area under curve	0.946	0.904	0.976	0.999	0.828	0.983	0.926	0.716	0.962
Asymptotic p-value	0.000	0.000	0.000	0.000	0.001	0.000	0.000	0.043	0.000
γcut-off	2.7	0.1	0.1	1.9	6.6	6.2	0.3	2.4	0.1
Sensitivity	0.879	0.849	0.930	1.000	0.784	0.965	0.869	0.709	0.924
Specificity	0.880	0.847	0.931	0.999	0.784	0.965	0.870	0.709	0.926
Fisher’s exact test p-value	0.000	0.000	0.007	0.000	0.018	0.000	0.002	0.281	0.000

^*Not available (N.A) cut-off threshold since five predictors were combined.

To demonstrate the potential utility of the improved DWI-MAP analysis in prediction of motor areas in clinical practice, Figure 4 shows an example of the results from improved DWI-MAP analysis; thereby, a 12-year-old girl (patient no. 14) developed hemiparesis requiring a rehabilitation therapy following surgical resection of the ECoG-defined seizure onset zone (red or green-colored spheres in left column). The DWI-MAP classifier delineated three motor pathways, C₁(face), C₂(hand) and C₃(leg) near the seizure onset zone, of which cortical terminals were spatially overlapped with ESM-defined motor areas for face, hand, and leg (blue or green-colored spheres in left column). Cortical terminals of C₁(face), C₂(hand), and C₃(leg) pathways were included in the surgically resected volume (blue colored area in right column). This illustrative case exhibits how the improved DWI-MAP analysis can be employed in presurgical evaluation of patients undergoing epilepsy surgery.

Figure 4.

Automatic detection of three primary motor pathways using the improved DWI-MAP classifier, C₁(face), C₂(hand), C₃(leg) obtained from patient no. 6. This patient has ECoG-determined seizure onset zone in the precentral gyrus (red or green-colored sphere in left column images). ESM electrode sites of face, finger, and leg motors (blue or green colored spheres) were spatially well-matched with cortical terminals of C₁(face), C₂(hand), C₃(leg) pathways, respectively. Cortical terminals of C₁(face), C₂(hand), and C₃(leg) pathways were included in surgically resected volume (blue colored area in right column images) causing severe motor impairment in this patient.

Discussion

The present study proposes to investigate the clinical utility of our novel DWI-MAP analysis^21,22 to improve the accuracy of localization of motor, language and visual areas for ultimately minimizing postoperative functional deficits in children undergoing brain surgery. ESM still remains to be the current gold standard for identifying motor, language and visual areas, and fMRI is used as a complementary tool in presurgical evaluation. Yet, ESM and fMRI fail to localize the ‘eloquent cortex’ of interest in 30–40% of pediatric cases^19,36. Furthermore, both ESM via subdural electrodes and hemodynamic responses on fMRI are inherently unable to localize crucial white matter pathways associated with eloquent cortex, and these pathways also may be at risk for collateral damage during epilepsy surgery. Thus, there is a need, especially in children, for identification of these cortical regions and their network connectivities during presurgical evaluation. We believe that a sophisticated DWI tractography analysis, as a non-invasive alternative modality, would satisfy this urgent need. Here, we improved the original DWI-MAP classifier^21,22 by adding a new clustering procedure using the pathway exemplars. The improved DWI-MAP approach indeed had a better performance in prediction of postoperative functional deficits. Our central hypothesis was that this novel DWI approach can serve as an accurate localizing tool in young children and can be used to estimate the occurrence of postoperative deficits in motor and language functions.

Two major findings emerge from the present study. First, by adding ADFD correction to the original DWI-MAP analysis, the proposed motor predictors Δ_1–3 and γ_1–3 could improve their sensitivity and specificity to predict postoperative deficits in the motor domain by 28% (from 0.62 to 0.79 on average) and 7% (from 0.81 to 0.87), respectively. In the prediction of postoperative language deficit using Δ_4–8 and γ_4–8, the addition of ADFD improved the sensitivity and specificity of our regression model by 13% (from 0.69 to 0.79) and 1% (from 0.86 to 0.87), respectively. In the prediction of postoperative visual deficit using Δ₉ and γ₉, the addition of ADFD improved the sensitivity and specificity of our regression model by 13% (from 0.77 to 0.87) and 24% (from 0.75 to 0.93), respectively. This suggests that ADFD correction could reduce misclassifications mainly attributable to equal a priori assumptions in the conventional Baysian framework^15,16. Secondly, preservation of the eloquent pathways defined by preoperative DWI-MAP with ADFD (i.e., γ_1–4,9) predicted the absence of postoperative deficits with sensitivity and specificity ranging from 88% to 100%. By co-registering surgically removed areas of postoperative MRI to preoperative MRI, these results show that up to 3% resection of DWI-MAP defined pathways (i.e., γ_cut-off) would provide the least likelihood of postoperative deficit in face motor, hand motor, leg motor, language, and visual function. With regard to the computation time, the improved DWI-MAP analysis takes about 30 minutes to identify 9 classes (C_1–9) from single DWI data using an in-house implementation combining Matlab (www.mathworks.com) and Python (www.python.org) at a graphical processing unit (NVidia GeForce GTX 1080 Ti).

Our findings suggest that postsurgical functional outcome may substantially differ according to the extent of resected white matter encompassing eloquent cortex determined by preperative DWI-MAP analysis, suggesting that DWI-MAP analysis can be a non-invasive imaging tool to help optimize surgical resection boundaries to minimize postoperative deficit. This could have important implications to preserve vital functions in surgical patients, particularly in young children whose ESM and fMRI studies sometimes fail to detect eloquent areas of interest. In fact, subgroup analyses reported in supplemental table 1 showed that up to 3% resection of DWI-MAP defined pathways (i.e., γ_cut-off) in children younger than 5 years old yielded the least likelihood of postoperative deficit in face motor (γ₁: TP = 4/5, FN=5/8), hand motor (γ₂: TP = 4/4, FN=9/9), leg motor (γ₃: TP = 4/4, FN=8/9), language (γ₄: TP = 3/3, FN=9/10), and visual field (γ₉: TP = 6/6, FN=5/7), supporting that the proposed DWI-MAP analysis can be an effective tool for surgical planning in young children. The improved DWI-MAP analysis can provide preoperative mapping of the patient’s “eloquent” cortical areas: C_1–9 without requiring additional modalities such as ESM and fMRI. Thus, the results of improved DWI-MAP analysis may be used for preoperative evaluation of patients who undergo single-stage resective epilepsy surgery. Many previous studies have attempted to classify the complicated tract patterns of whole-brain DWI tractography by objectively recognizing their shape, length, and anatomical features in the frameworks of virtual dissection, based on anatomical information,^7,8,24 supervised clustering using similarity metrics,^10,13,14,32 atlas based annotation,^17,34,41 and deep learning network²⁶. These approaches have developed to achieve the same objectives: removing superficial tracts in whole brain tractography and saving effort in white matter dissection by accurately extracting anatomically known white matter bundles using a set of prior models. In contrast to other approaches, the present study is focused on clinical validation of automatic white matter dissection by performing systemic comparisons of ADFD-corrected DWI-MAP analyses to detect electrophysiologically-confirmed eloquent white matter pathways in children.

The present study does have at least two limitations. First, all data were obtained for the clinical management of epilepsy surgery. Therefore, the findings of this study are naturally based on a retrospective and observational design. Such design inherently limits the sample sizes of patients with postoperative deficit, restricting availability of detailed functional testing and the measures that are dependent on that testing: namely the interaction of γ_1–9 with biological and clinical variables (see the results of different subgroup analyses at age and pathology in supplemental table 1). Further studies and replication are needed in larger numbers of children with a greater variety of postoperative deficits and deficit severity. Also, it is worth noting that the main goal of this study was to explore whether preservation of the eloquent pathways defined by the improved DWI-MAP analysis was associated with successful avoidance of acute, newly-developed functional deficits in different somatosensory, motor, language and visual domains. Thus, long-term, quantitative, longitudinal assessment of change in postoperative deficit is beyond the scope of this study. Future studies are warranted to determine the temporal change in tract volume as a function of functional improvement. Comparative studies have to be performed to investigate whether preservation of the eloquent pathways defined by the improved DWI-MAP analysis is predictive of both short- and long-term functional outcomes. In addition, both head size and extent of seizure focus may provide different degree of freedom in registering stereotaxic probability maps of eloquent white matter pathways from template space to native space, which may degrade the predictability of the proposed DWI-MAP marker: γ_cut-off. These two confounders should be carefully taken into account to apply the presented values of γ_cut-off (e.g., Table 4) to predict postoperative deficits in younger children.

Second, it remains controversial that DWI can accurately derive long-range anatomical connections,³⁸ especially in current clinical settings where the limited scan time allows only a small number of diffusion encoding data measured at a low diffusion factor (i.e., b-value). Moreover, despite the exceptional quality of DWI data using high angular resolution diffusion imaging (HARDI), a recent ex-vivo macaque DWI study demonstrated that local fiber orientation data has inherent limitations when determining long-range anatomical projections, such as those connecting the precentral gyrus to the rostroventral part of the occipital cortex.³⁸ Also, another ex-vivo macaque study with histological analysis reported that about 50% of the cortical surface (most often from sulcal regions but also from many gyral crowns) is inaccessible for long-range diffusion tracking due to the presence of dense white matter zones beneath the infragranular layers of the cortex, suggesting a new tractography paradigm needed to successfully perform tracking from superficial white matter into long-range anatomical connections accurately.³⁵ Considering both the technological and inherent ambiguities of current DWI tractography, the scope of this study was limited to consider only 1) short-range motor pathways connecting primary motor cortex to the posterior limb of the internal capsule and 2) major pathways of language and visual segments, which are anatomically consistent with dissection studies.^12,39 Thus, the proposed DWI-MAP analysis is limited in its investigation of primary motor and language-visual system response to surgical intervention, since it cannot reliably explore the long-range cortical and subcortical connections within these systems.⁴⁰

Nevertheless, the current findings do provide new support for the utility of DWI tractography in pediatric epilepsy surgery. Although advances in neuroimaging technology have improved mapping of eloquent areas in the last decades, studies have reported that postoperative neurological deficits range from 50% to 83% in patients from the conventional resection cohort,^5,6,33 and even after intraoperative-MRI-assisted resection, 36% of patients still show postoperative neurological deficits.³⁷ With this in mind, the present study aimed to improve a clinically effective imaging method by ADFD-correcting DWI-MAP analysis, facilitating objective and sustainable detection of functionally important white matter pathways without other supplementary modalities. Systematic and prospective investigation of the proposed method will further improve presurgical planning and provide a unique opportunity to minimize or predict postsurgical functional deficits in the future.