Inclusion criteria were (1) PFT patients treated at Gustave Roussy Cancer Campus between 2000 and 2014; (2) 17 years of age or under at diagnosis (3) multiple (>2) IQ assessments after treatment onset (4); for the PFT patients treated with radiation therapy, the computed tomography (CT) scan, T1-weighted MRI, and dosimetric maps had to be available. Thirty patients (14 males and 16 females) matched these criteria. Information was gathered from medical files about the history of the illness (i.e., age at diagnosis) and the type of treatment (i.e., surgery, chemotherapy, and radiation therapy protocol). The underlying malignancy of the 30 patients studied was medulloblastoma, ependymoma, astrocytoma, and embryonal tumor in 25, 3, 1, and 1 patients, respectively. Twenty patients had a localized disease and 10 had a metastatic disease. Complete tumor resection was achieved in 20 PFT. Post-operative complications occurred in 10 patients. No patient relapsed between the two evaluations, but the patient with an astrocytoma whose relapse before the first evaluation, was treated with chemotherapy alone. The mean age at diagnosis was 4.62 years (SD = 3.05; [0.49; 12.24]). The mean delay between treatment and the last assessment was 4.60 years (SD = 4.60; [1.28; 14.24]). Pre-operative hydrocephalus was present in 19 patients (63%). Seventeen patients were treated with RT alone (N = 7) or RT and chemotherapy (N = 10). The remaining patients were treated with chemotherapy alone and were used as controls. All patients with RT received a CSI and a boost in the PF and were treated with three-dimensional conformal radiation therapy (Table 1). This study was approved by an ethical committee (CPP no. 14973, Gustave Roussy, Villejuif, France).

Three cognitive indices were estimated from age appropriate Wechsler Intelligence Scale (20, 21): FSIQ in all patients and the processing speed index (PSI) and working memory index (WMI) when available. Neuropsychological assessments were done by formally trained neuropsychologists from the pediatric department. Because of age or time constraints, not all participants were administered all the tests. Thus, the numbers of patients assessed varied across cognitive scores [N (ΔFSIQ) = 30, N (ΔPSI) = 23, N (ΔWMI) = 14]. Patients were evaluated at variable time points after treatment onset. Thus, the delay (T) between two neuropsychological assessments varied from one patient to the other (Table 2). The change in cognitive scores (ΔFSIQ, ΔPSI, and ΔWMI) of each patient was calculated from the difference between the first and last scores (ΔT). We did not consider intermediate scores. Changes in cognitive scores (ΔFSIQ, ΔWMI, and ΔPSI) were compared using two-tailed Student’s t-tests.

3D T1-MR images were acquired on a 3-T scanner using a 32-channel head coil (General Electric, Milwankee, MN, USA). In this clinical retrospective study, two types of T1-weighted images were collected: 3D T1-weighted sagittal slices (matrices: 256 mm × 256 mm, pixel size: 0.5 mm, slice thickness: 1 mm, FOV: 240 mm) and 3D-T1 weighted axial slices (matrices: 224 × 288, pixel size: 0.5 mm, slice thickness: 1 mm).

Computed tomography scans were acquired on a SIEMENS Sensation Open scanner located in Gustave Roussy RT department (matrices: 512 mm × 512 mm, pixel size: 0.8 mm, slice thickness: 3 mm). Radiation dose maps (RD maps) were computed with the ISOgray™ Treatment Planning System (DOSIsoft, version 4.1, Cachan, France). The Clarkson–Cunningham model was used for dose calculation. Dose maps resolution was 3 mm × 3 mm × 1 mm.

At time of first neuropsychological assessment, the mean estimated IQ over the whole population was 87.5 (SD = 18.4; [45–130]). A declining performance over time was observed in 67, 64, and 48% of the patients for ΔFSIQ, ΔWMI, and ΔPSI, respectively. The remaining patients showed either preserved or better performance over time. However, there were no statistically significant differences between cognitive scores [ΔFSIQ vs. ΔWMI: t(32) = −0.64, p = 0.52; ΔPSI vs. ΔWMI: t(35) = −0.81, p = 0.42; ΔFSIQ vs. ΔPSI: t(51) = −0.37, p = 0.70] (Table 2). Moreover, ANOVAs were conducted to compare the three treatment groups (chemotherapy alone vs. RT and chemotherapy vs. RT alone) on their cognitive scores (ΔFSIQ, ΔWMI, and ΔPSI). There was no statistically significant difference between treatment groups in ΔFSIQ [F(2,30) = 2.36; p = 0.11; RT alone: M = −10.5 (±9.11), chemotherapy alone: M = 1.57 (±11.28), RT and chemotherapy: M = −2.0 (±11.03)]; and ΔWMI [F(2,14) = 1.17; p = 0.34; RT alone: M = −10 (±9.90), chemotherapy alone: M = −0.29 (±5.90), RT and chemotherapy: M = −4.60 (±10.15)], and ΔPSI [F(2,24) = 2.28; p = 0.12; RT alone: M = −10 (±9.90), chemotherapy alone: M = −0.29 (±5.90), RT and chemotherapy: M = −4.60 (±10.15)].

PC1-_EUD, which explained 67% of the variance of original data, was strongly correlated (>0.50) with the dose (EUD) in all regions, especially in the supratentorial space. PC2-_EUD explained 19% of the variance and was positively correlated with 16 regions in the PF, inferior occipital and temporal regions (see Table S1 in Supplementary Material). Meanwhile, three regions in the left superior occipital and parietal regions correlated negatively and moderately (>0.40) with PC2-_EUD (see Table S1 in Supplementary Material). PC3-_EUD explained 5% of the variance and had a moderate positive correlation (>0.40) with the EUD in the left orbitofrontal area. By contrast, the precuneus and the right cuneus negatively correlated with PC3-_EUD. Values of all correlation coefficients are shown in Table S2 in Supplementary Material.

Our final regression models included the three PCs-_EUD, chemotherapy, age at diagnosis, and delay between assessments (ΔT). Indeed, in all models, time since diagnosis (variance inflation factor >10) was highly correlated with ΔT (variance inflation factor >10), while it was not the case between PCs (variance inflation factor <10), chemotherapy (variance inflation factor <10), and age at diagnosis (variance inflation factor <10). We thus removed time since diagnosis from the analysis and checked that all remaining variance inflation factor indices were below 10.

Consistently with previous studies (5, 6), the FSIQ decline depended on the delay between the two IQ tests. As shown in previous studies (5, 38), chemotherapy does not seem to have a significant negative impact on PSI and WMI functioning. The surprising positive effect of chemotherapy on FSIQ change might be linked to the positive impact of repeated measurements, also known as the carry over effect (or IQ test–retest) (39). Children acquired expertise concerning neuropsychological task along many neuropsychological tests, improving their performances. Therefore, the change in cognitive scores of each patient calculated from the difference between the first and last scores was positive. A large portion of children with chemotherapy alone showed an IQ improvement which confirms the absence of cognitive effect of chemotherapy (5, 38).

This also explains the unexpected negative effect of age at diagnosis on FSIQ change, as children treated with chemotherapy alone are usually young (below 5 years) at diagnosis.

All components seem to have specific impacts on changes of the working memory score (WMI). The radiation distribution pattern involving the left orbitofrontal regions (PC3-_EUD) had the most negative impact on working memory. Interestingly, this result could be in line with Mabbott et al.?’s findings (40). They observed that working memory performance over time was different according to the tumor location in children treated for a central nervous system germ cell tumor. Patients with pineal tumors showed early, but stable, working memory deficit, whereas patients with suprasellar tumors experienced a significant working memory decline over time. Mabbott et al. suggested the observed decline was related to the radiation field rather than to the tumor location (40). In addition, this observation fits well with the compelling neuroimaging evidence of orbitofrontal implication in tasks relying on working memory [for meta-analysis, see Ref. (41, 42)]. PC1-_EUD, however, corresponds to a distributed radiation pattern across the whole brain, suggesting that a global increase of radiation dose (EUD) impacts working memory negatively. From its patterns of spatial radiation distribution, this last component could be interpreted as CSI dose variability across subjects. However, such an overall radiation effect does not allow us to distinguish specifically irradiated brain networks that could be particularly involved in working memory impairment.

More specific brain network radiation patterns are found to influence processing speed. The large impact of PC3-_EUD is strongly related to radiation to the temporal lobes and the PF. Previous studies have shown significant associations between radiation dose to the temporal lobe and processing speed impairments (16, 17, 43). The cerebellum has also been shown to play a role in processing speed capabilities (44). Importantly, temporal lobe regions are close to the PF upon which the dose was escalated. Thus, PC3-_EUD impact could also reflect the radiation field boost trajectory to the PF across subjects. This would support the hypothesis that the volume receiving the highest dose has the greatest impact on cognitive functions. Accordingly, these findings would support current volume–reduction efforts.

Finally, PC2-_EUD and PC3-_EUD carry the exact same negative effect on IQ change. As reported earlier, PC2-_EUD that includes the temporal lobes and the PF showed the most significant impact on processing speed changes. As for processing speed, previous studies have found associations between radiation dose to the temporal lobe and PF and IQ impairments (16, 18). In the same way, the role on IQ impairment of PC3-_EUD, which strongly involves the left orbitofrontal cortex, is somehow expected, as many VBM studies in adults and adolescents have shown a link between IQ and gray matter density in this region (45–47). Alternatively, the equal contribution of these two components on ΔFSIQ might be the expression of an averaging effect as FSIQ is a composite index encompassing both working memory and processing speed.

Higher EUD in the superior occipital and parietal regions did not seem to be associated with lower cognitive scores. We may note that Armstrong et al. (17) did not find any significant association between occipito-parietal radiation dose and cognitive or social problems either. In addition, with the same amount of radiation dose, the parietal lobe white matter was shown to be less affected compared to frontal lobe in medulloblastoma (48, 49). Thus, it would be interesting to test whether the parietal lobe is less susceptible to radiation than other regions.

There are limitations in this study, and results should be interpreted with caution. First, the small size and heterogeneity of the patient population make it difficult to control for other variables that could affect the scores (i.e., hydrocephalus shunt, education, rehabilitation, surgical approach, molecular group, etc.). Moreover, considering only PFT patients prevented us from taking into account several potentially confounding variables such as type and localization of the tumor. However, this was a disadvantage regarding the large spatial correlations between close irradiated regions induced by similar radiation protocol. We could not access separately specific regions that are known to play an important role in working memory [e.g., dorsolateral area (50)] or processing speed [e.g., left middle frontal gyrus (51)]. Second, noise could be induced by intersubject variability of the brain morphology, even if we minimized possible segmentation errors by using atlases specific to age groups. Finally, we have to acknowledge that seven patients (that received hyperfractionated radiotherapy, HFRT) had the same total dose and could be considered as a subgroup that could influence the results (see Figure S1 in Supplementary Material). We recognize the possibility that the HFRT subgroup smaller variance might influence the result in other less crucial ways (see Figure S2 in Supplementary Material).