Synopsis

Introduction: Posterior fossa tumors (Medulloblastoma, Ependymoma, Pilocytic astrocytoma, and uncommon ATRT) are notoriously difficult to differentiate on pre-operative MRI scans. Despite using advanced MRI sequences (diffusion, perfusion weighted, etc. ) expert interpreters are at best 50-60% accurate in predicting the different histologies. Voxel level tumoral histogram analytics has been recently used to grade GBM, predict response to therapy and even tumor subtyping. We explored the utility of lesional histogram metrics derived from multiparametic MRI (T1C, ADC maps), in preoperative differentiation.

METHODS: Retrospective review of pathology archives form 2007 to current reveled 160 pts. with histologically proven posterior fossa tumors, 52/160 pts. (Male = 30 , female = 22 , mean age = 7.8 ( range from 6 months to 18 year) ; medulloblastoma 18, ependymoma 12, pilocytic astrocytoma 17 , ATRT 5 ) who underwent routine preoperative MRI scanning, and restricted to pts. undergoing acquisition on GE 1-5T (GE Signa HDxt, Waukesha, WI, USA) were selected for further analysis. All MRI were performed per accepted protocols, the ADC maps derived from DWI, T-2 and FLAIR images were coregistered to TI contrast enhanced datasets on a dedicated review workstation (MIMsoftware, Cleveland, OH, USA). All tumors, and contralateral normal brain were manually flagged in 3D using the T1C along with T-2/FLAIR, the VOI’s were synchronously copied to all open coregistered datasets. The voxel level data from ADC maps and T1C were exported to EXCEL worksheet. Tumoral ADC/perfusion values were normalized to contralateral normal brain. Subsequently, histogram metrics (mean, standard deviation, skewness, Kurtosis, and percentiles, etc.) were computed from the normalized estimates, using a home built EXCEL macro. ANOVA was performed for all histogram metrics against the commonly occurring tumor histologies. Furthermore, all derived histogram metrics (features) were data mined with histology as class (ground truth) for prospective histologic prediction, using freely available software WEKA (Waikato Univ. New Zealand).

Results: On ANOVA analysis only normalized ADC-kurtosis reached statistical significance (p-value – 0.04) for differentiation, while the ADC-Minimum (p value 0.073) was significant at 0.10 level, although not reaching the conventional alpha level 0.05. Data mining (in 49/52 pts) the histogram metrics as features for histologic classification and prediction, among various classification algorithms in WEKA, the algorithm J48 was the best predictor (using 10 fold leave one out, for training and testing) with a combined perfusion and ADC feature subset selection (ADC_Mean, PERF_Kurtosis, PERF_Skewness, PERF_Minimum, and ADC_Maximum) correctly predicting 28/41 (57%), and incorrect prediction 21/41 (43%), for a ROC of 0.76. However, when using subset of features derived from ADC alone or perfusion alone, the correct and incorrect classification was 16/49 and 33/49 for ROC of 0.57 with Naïve Bayes, and 17/49 and 32/49 again again with Naïve Bayes for a ROC of 0.76.

Discussion: Voxel level histogram analytics and data mining of the histogram features in our series corroborates recent work by Redriguez et al (1) and others (2-4). While Redriguez and co-workers utilized only metrics derived from a single sequence - ADC alone, majority of other groups utilized only single representative slices in their analysis. However, in our series we performed 3D manual segmentation of tumors, in addition, we utilized two functional MRI sequences perfusion and diffusion. Though, the results of above groups are superior and impressive; their results may be due to choice of representative slices, rather than the whole tumor analysis, and also not including the uncommon ATRT in their analysis. In our series, utilizing the subset of features derived from both ADC and perf turned out to be most accurate.

Conclusion: The findings in our series suggests that 3D voxel level histogram analysis of multiparametic FMRI (perfusion and diffusion) and data mining techniques is possible and quite promising for preoperative differentiation of posterior fossa tumors. We believe, we are the 1st group to perform multiparametric MRI histogram analytics of this entity in 3D, Future analysis with additional sequences is anticipated.

Acknowledgements

References

References: 1. Metrics and Textural Features of MRI Diffusion to Improve classification of Pediatric Posterior Fossa Tumors. D. Rodriguez Gutierrez et al, AJNR Am J Neuroradiol 2014, pg 1-7

2. Improving tumour heterogeneity MRI assessment with histograms. N Just et al, Br J Cancer. 2014 Dec 9; 111(12): 2205–2213.

3. Apparent Diffusion Coefficients for Differentiation of Cerebellar Tumors in Children. Z. Rumboldt et al, JNR Am J Neuroradiol 27:1362–69 Jun-Jul 2006.

4. Discrimination of paediatric brain tumours using apparent diffusion coefficient histograms. Jonathan G. Bull et al, Eur Radiol (2012) 22:447 – 457.