Brain tumour cell segmentation and detection using deep learning networks. Issue 10 (23rd April 2021)

• Record Type: Journal Article
• Authors: Bagyaraj, Sanjeevirayar; Tamilselvi, Rajendran; Mohamed Gani, Parisa Beham; Sabarinathan, Devanathan
• Is Part Of: IET image processing. Volume 15:Issue 10(2021)
• Page Start: 2363
• Page End: 2371
• Publication Date: 2021-04-23
• Subjects: Image processing -- Periodicals; 621.36705
• DOI: 10.1049/ipr2.12219
• Languages: English
• ISSNs: 1751-9659

Abstract

Abstract: Medical science is a challenging area for various problems associated with health care and there always exists scope for continuous medical research. The major challenges in medical imaging are in the region of lesion, segmentation and classification of tumours in the brain. Several technical challenge exists in the classification due to the variation in the tumour size, shape, texture information and location. There is a need for automatic identification of high‐grade glioma (HGG) and lower‐grade glioma (LGG). The management and grade of brain tumour depend on the depth of the tumour. Due to its irregular features, manual segmentation involves longer time and also increases the misclassification rate. Inspired by these issues, this paper introduces two automatic deep learning networks called U‐Net‐based deep convolution network and U‐Net with dense network. The proposed method is evaluated in our own brain tumour image database consisting of 300 high‐grade brain tumour cases and 200 normal cases. To improve the overall efficiency of the network, data augmentation is applied in both training and validation. The proposed U‐Net‐based Dense Convolutional Network (DenseNet) architecture is compared with the performance of U‐Net architecture and concluded that the proposed DenseNet produces a higher dice value. The validation results have revealed that our proposed method can have better segmentation efficiency. Also, the performance of the proposed DenseNet achieved better results compared with the state‐of‐the‐art algorithms. Validation of the test images proves that segmented output classification of tumour risk and the normal region produces a sensitivity of 88.7%, Jaccard index of 0.839, dice score value of 0.911, F1 score of 0.906 and specificity of 100% using U‐Net‐based DenseNet architecture.