Multi-stream Convolutional Autoencoder and 2D Generative Adversarial Network for Glioma Classification
Paper i 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.

Författare

Muhaddisa Barat Ali

Chalmers, Elektroteknik, Signalbehandling och medicinsk teknik, Signalbehandling

Irene Yu-Hua Gu

Chalmers, Elektroteknik

Asgeir Store Jakola

Sahlgrenska universitetssjukhuset

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

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

Ämneskategorier

Systemvetenskap

Datavetenskap (datalogi)

Datorsystem

DOI

10.1007/978-3-030-29888-3_19

Mer information

Senast uppdaterat

2019-11-06