Multiclass magnetic resonance imaging brain tumor classification using artificial intelligence paradigm

Tandel G. S.;Balestrieri A.;Jujaray T.;Khanna N. N.;Saba L.

Penultimate

;Suri J. S.

2020-01-01

---

Abstract

Motivation: Brain or central nervous system cancer is the tenth leading cause of death in men and women. Even though brain tumour is not considered as the primary cause of mortality worldwide, 40% of other types of cancer (such as lung or breast cancers) are transformed into brain tumours due to metastasis. Although the biopsy is considered as the gold standard for cancer diagnosis, it poses several challenges such as low sensitivity/specificity, risk during the biopsy procedure, and relatively long waiting times for the biopsy results. Due to an increase in the sheer volume of patients with brain tumours, there is a need for a non-invasive, automatic computer-aided diagnosis tool that can automatically diagnose and estimate the grade of a tumour accurately within a few seconds. Method: Five clinically relevant multiclass datasets (two-, three-, four-, five-, and six-class) were designed. A transfer-learning-based Artificial Intelligence paradigm using a Convolutional Neural Network (CCN) was proposed and led to higher performance in brain tumour grading/classification using magnetic resonance imaging (MRI) data. We benchmarked the transfer-learning-based CNN model against six different machine learning (ML) classification methods, namely Decision Tree, Linear Discrimination, Naive Bayes, Support Vector Machine, K-nearest neighbour, and Ensemble. Results: The CNN-based deep learning (DL) model outperforms the six types of ML models when considering five types of multiclass tumour datasets. These five types of data are two-, three-, four-, five, and six-class. The CNN-based AlexNet transfer learning system yielded mean accuracies derived from three kinds of cross-validation protocols (K2, K5, and K10) of 100, 95.97, 96.65, 87.14, and 93.74%, respectively. The mean areas under the curve of DL and ML were found to be 0.99 and 0.87, respectively, for p < 0.0001, and DL showed a 12.12% improvement over ML. Multiclass datasets were benchmarked against the TT protocol (where training and testing samples are the same). The optimal model was validated using a statistical method of a tumour separation index and verified on synthetic data consisting of eight classes. Conclusion: The transfer-learning-based AI system is useful in multiclass brain tumour grading and shows better performance than ML systems.