Get access

A dual neural network ensemble approach for multiclass brain tumor classification

Authors


  • The work has been conducted as a collaborative project to develop an interactive CAD system to assist radiologists under MOU between IIT Roorkee, India and PGIMER, Chandigarh, India

Jainy Sachdeva, Electrical Engineering, Indian Institute of Technology Roorkee, Roorkee, India.

E-mail: jainysachdeva@gmail.com

SUMMARY

The present study is conducted to develop an interactive computer aided diagnosis (CAD) system for assisting radiologists in multiclass classification of brain tumors. In this paper, primary brain tumors such as astrocytoma, glioblastoma multiforme, childhood tumor-medulloblastoma, meningioma and secondary tumor-metastases along with normal regions are classified by a dual level neural network ensemble. Two hundred eighteen texture and intensity features are extracted from 856 segmented regions of interest (SROIs) and are taken as input. PCA is used for reduction of dimensionality of the feature space. The study is performed on a diversified dataset of 428 post contrast T1-weighted magnetic resonance images of 55 patients. Two sets of experiments are performed. In the first experiment, random selection is used which may allow SROIs from the same patient having similar characteristics to appear in both training and testing simultaneously. In the second experiment, not even a single SROI from the same patient is common during training and testing. In the first experiment, it is observed that the dual level neural network ensemble has enhanced the overall accuracy to 95.85% compared with 91.97% of single level artificial neural network. The proposed method delivers high accuracy for each class. The accuracy obtained for each class is: astrocytoma 96.29%, glioblastoma multiforme 96.15%, childhood tumor-medulloblastoma 90%, meningioma 93.00%, secondary tumor-metastases 96.67% and normal regions 97.41%. This study reveals that dual level neural network ensemble provides better results than the single level artificial neural network. In the second experiment, overall classification accuracy of 90.4% was achieved. The generalization ability of this approach can be tested by analyzing larger datasets. The extensive training will also further improve the performance of the proposed dual network ensemble. Quantitative results obtained from the proposed method will assist the radiologist in forming a better decision for classifying brain tumors. Copyright © 2012 John Wiley & Sons, Ltd.

Ancillary