Multi-stream Convolutional Autoencoder and 2D Generative Adversarial Network for Glioma Classification
Paper in proceeding, 2019

Diagnosis and timely treatment play an important role in preventing brain tumor growth. Deep learning methods have gained much attention lately. Obtaining a large amount of annotated medical data remains a challenging issue. Furthermore, high dimensional features of brain images could lead to over-fitting. In this paper, we address the above issues. Firstly, we propose an architecture for Generative Adversarial Networks to generate good quality synthetic 2D MRIs from multi-modality MRIs (T1 contrast-enhanced, T2, FLAIR). Secondly, we propose a deep learning scheme based on 3-streams of Convolutional Autoencoders (CAEs) followed by sensor information fusion. The rational behind using CAEs is that it may improve glioma classification performance (as comparing with conventional CNNs), since CAEs offer noise robustness and also efficient feature reduction hence possibly reduce the over-fitting. A two-round training strategy is also applied by pre-training on GAN augmented synthetic MRIs followed by refined-training on original MRIs. Experiments on BraTS 2017 dataset have demonstrated the effectiveness of the proposed scheme (test accuracy 92.04%). Comparison with several exiting schemes has provided further support to the proposed scheme.


Muhaddisa Barat Ali

Chalmers, Electrical Engineering, Signal Processing and Biomedical Engineering

Irene Yu-Hua Gu

Chalmers, Electrical Engineering

Asgeir Store Jakola

Sahlgrenska University Hospital

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

03029743 (ISSN) 16113349 (eISSN)

Vol. 11678 LNCS 234-245
978-3-030-29888-3 (ISBN)

International Conference on Computer Analysis of Images and Patterns
Salerno, Italy,

Subject Categories

Information Science

Computer Science

Computer Systems



More information

Latest update