Synopsis

Introduction: Multiparametric MRI is only modestly accurate for pre-operative diagnosis of posterior fossa tumors by domain experts (experienced radiologists), the addition of even advanced FMRI (advMRI) techniques (perfusion, diffusion, DTI), has not resulted in impressive accuracy. The emerging field of RADIOMICS which entails high throughput extraction of hidden features from medical image data sets, has been notably successful in oncologic imaging, for example in tumor staging/grading, response to therapy, and clinical outcomes, in various intra and extra cranial tumors, including GBM. We postulate that higher order hidden textural features and data mining the extracted features can significantly improve the preoperative predictive accuracy of posterior fossa tumors, as compared to domain experts..

Methods: Retrospective inter-observer (completed) and intra-observer variation (in-progress, by three experienced neuroradiologists (MUM-consultant, ALD and ALR –neuroradiology fellows) was performed in a subset of 52/160 pts, with histological diagnosis of posterior fossa tumors from 2007-current, and also having undergone preoperative MRI in our institution on 1.5T magnet (GE Signa HDxt, Waukesha, WI, USA.). The multiparametric MRI sequences (T1C, T2, Flair and ADC maps) were all coregistered to a common T1C-3D image space, manual 3D VOI’s of lesions were co-flagged using all the sequences for guidance, synchronously on all open datasets, utilizing a dedicated workstation (MIMsoftware, Cleveland, OH, USA) . Subsequently, a subset of 13/52 pts (medulloblastoma=7, pilocytic astrocytoma=4, ependymoma=2) was chosen for the pilot textural analysis. The tagged advMRI were pipelined through MaZda ( public domain texture analysis suite) to extract various textural features (no.=110). Subsequently, the textural feature output from MaZda were data mined with known histological classes, to build classification and predictive models in Weka (data mining suite, Univ. of Waikato, New Zealand). The best feature subset selection/s derived from ADC alone, perfusion alone, and merged ADC and perfusion textural features were utilized, for training, and testing a number of predictive models, separately for each of the above feature subset. The best predictive models were subsequently validated (10 fold leave one out) and compared with domain experts.

Results: The inter observer variability between the three experts was substantial. The kappa statistic between MUM and ALD; MUM and ALR; and lastly ALD and ALR were 0.46, 0.49 and 0.39 with a mean of 0.45. We evaluated different classifier models, and the most predictive for the three textural subsets derived from ADC alone, perf alone, and lastly the merged ADC+perf feature set were found to be the classifiers J48 and Naive Bayes . Validation of the most predictive models (10 fold leave one out) were equivalent, for the above three subsets – weighted avg. ROC of 0.75 to 0.76, all the top performing models (ADC, perf, and merged) correctly classified 7/13 instances (56%), while misclassifing 6/13 ( 46%).

. . Discussion: As expected the inter-observer variation was substantial. ;As opposed to recent work by others who employed a single sequence – ADC for prediction (1,3,4) we utilized data from multiple sequences for developing predictive models with the extracted textures. Furthermore, as opposed to findings in (1-4) in our series perfusion textural features as well as the merged features performed equivalent to the ADC derived classifier models. Though, in our study the prediction/classification by all top-performing models including ADC were not as high as reported recently 85-95th percentiles with ADC (1,3,4) Vs. 75th percentiles in our series; the predictive rates may be explained on basis of 1. Small sample size 2. Unequal distribution of samples 3. Potential susceptibility artifacts corrupting ADC datasets.

Limitations: 1. The sample size is quite small for the large feature dimensionality 2. Some of the ADC image data may be potentially corrupted due to know susceptibility artifacts at base of brain and tissue interphases. 3. Unequal distribution of histopathology

Conclusion: The observations in our pilot series corroborates emerging evidence, that perfusion textural analysis can be of significant utility in preoperative differentiation of posterior fossa tumors, as authenticated in various tumors, including GBM. Furthermore, our study has uniquely utilized multisequence MRI texture analytics and data mining for preoperative prediction/classification of these lesions. The pilot analysis suggests that RADIOMICS techniques may eventually be equal to or better than domain experts in this entity. We believe that we may be the 1st to demonstrate the utility of perfusion textures in differentiating these uncommon but clinically challenging tumors.

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.