Table 1 CNN studies and main findings obtained from literature search
From: Convolutional neural networks for brain tumour segmentation
Author, year | Title | Main findings |
|---|---|---|
Pereira, 2016 [30] | Brain Tumour Segmentation Using Convolutional Neural Networks in MRI Images | Small 3 × 3 kernels for convolution to combat overfitting. DSC—complete tumour, 0.78; core tumour, 0.65; and enhancing regions, 0.75 |
Arunachalam, 2017 [23] | An efficient and automatic glioblastoma brain tumor detection using shift-invariant shearlet transform and neural networks | Unique segmentation process involving SIST and NSCT transformation to convert the image into a multi-resolution image. Standard feature extraction occurs. Accuracy is reported at 99.8% for the proposed method |
Havaei, 2017 [28] | Brain tumour segmentation with Deep Neural Networks | The TwoPathCNN (focusing on local and global paths) resulted in a DSC of complete segmentation, 0.85; core, 0.78; and enhancing, 0.73 |
AlBadawy, 2018 [31] | Deep learning for segmentation of brain tumours: Impact of cross-institutional training and testing | Training data on different institutions may produce dramatically different results. Therefore, CNNs need to be trained on data from the same institution |
Hasan, 2018 [24] | A Modified U-Net Convolutional Network Featuring a Nearest-neighbour Re-sampling-based Elastic-Transformation for Brain Tissue Characterization and Segmentation | Traditional U-net ‘deconvolves’ the voxels rather than convolving. For this study, the deconvolution layer is substituted with an upsampling layer which passes through two convolution layers, an upsampling layer followed by augmentation by elastic transformation. DSC increased from 0.86 to 0.87. |
Naceur, 2018 [29] | Fully Automatic Brain Tumour Segmentation using End-To-End Incremental Deep Neural Networks in MRI images | Incremental technique based on DSC which ‘learns’ features of scan until no features are learnt that increase DSC. This is iteratively refined. The DSC for this model is whole tumour, 0.89; tumour core, 0.76; and enhanced tumour, 0.81 |
Perkuhn, 2018 [22] | Clinical Evaluation of a Multiparametric Deep Learning Model for Glioblastoma Segmentation Using Heterogeneous Magnetic Resonance Imaging Data from Clinical Routine | Evaluation of DeepMedic architecture. DSC—whole tumour, 0.86; contrast enhanced tumour, 0.78; and necrosis, 0.62 |
Chang, 2019 [27] | A mix-pooling CNN architecture with FCRF for brain tumour segmentation | For global context, a fully connected conditional random field was combined to the CNN. DSC of complete tumour, 0.80; core tumour, 0.75; and enhancing, 0.71 |
Sundararajan, 2019 [25] | Convolutional Neural Network Based Medical Image Classifier | Segmentation by the watershed algorithm as opposed to manual segmentation for training sets improved accuracy from 82 to 89%. DSC is not reported |
Chang 2019 [19] | Automatic assessment of glioma burden: a deep learning algorithm for fully automated volumetric and bi-dimensional measurement | Skull-stripping was superior to other methods proposed in the literature. FLAIR hyperintensities relating to oedema were able to be delineated in a multi-institutional context, pre- and post-operatively. DSC was 0.917 for the FLAIR volume |