Brain tumour classification using siamese neural network and neighbourhood analysis in embedded feature space. Issue 3 (20th January 2021)

• Record Type: Journal Article
• Authors: Deepak, S.; Ameer, P. M.
• Is Part Of: International journal of imaging systems and technology. Volume 31:Issue 3(2021)
• Page Start: 1655
• Page End: 1669
• Publication Date: 2021-01-20
• Subjects: brain tumour -- classification -- Mahalanobis distance -- neighbourhood -- Siamese networks; Imaging systems -- Periodicals; Image processing -- Periodicals; 621.367
• DOI: 10.1002/ima.22543
• Languages: English
• ISSNs: 0899-9457

Abstract

Abstract: The application of deep transfer learning techniques has been successful in developing accurate systems for brain tumour classification on large‐scale medical image databases. For small databases, feature learning by deep neural networks is not robust. The systems based on domain‐specific hand‐crafted features have limited accuracy. In this paper, the authors focus on developing accurate models that could be trained effectively using a smaller number of data samples. A siamese neural network (SNN) is designed to extract features from brain magnetic resonance imaging (MRI) images. The SNN is realised using a 3‐layer, fully connected neural network. The designed SNN has lesser complexity and fewer parameters than deep transfer‐learned convolutional neural networks (CNN). A nearest neighbourhood analysis, using Euclidean and Mahalanobis distances, is conducted on the SNN encoded feature space. The encoded feature space is two dimensional, such that the neighbourhood analysis is computationally less intensive. For the neighbourhood analysis, a k‐nearest neighbour (k‐NN) model is utilised. The proposed method is evaluated using three publicly available datasets, namely, Radiopaedia, Harvard and Figshare repositories. The respective classification accuracy on cross‐validation is 92.6%, 98.5% and 92.6%. Other metrics used for the performance evaluation include F‐score, Specificity and balanced accuracy. The underlying network architecture and the design choice of network layers allow the implementation of the SNN in environments with low computational resources. The SNN features are found to be more effective than the hand‐designed features, and the deep transfer learned features for the stated problem.