Automatic segmentation of brain tumour in MR images using an enhanced deep learning approach

The presented manuscript proposes a fully automatic deep learning method to quantify the tumour region in brain Magnetic Resonance images as the accurate diagnosis of brain tumour region is necessary for the treatment of the patients. The irregular and confusing boundaries of tumours regions make it a challenging task to accurately figure out such regions. Another challenge with the segmentation task is of preserving the boundary details of the segmented tumour regions. The proposed network focuses on delineating the irregular tumour region as the best feature maps are learnt by the network, which is used for decoding; thus, it preserves the accurate boundary and pixel details. The proposed method incorporates internal residual connections in encoder and decoder to transfer feature maps directly to the successive layers to avoid loss of information contained in the images. The use of cross channel normalization (CCN) and parametric rectified linear unit (PRELU) gives a more balanced network output. The trained network produced remarkable results when tested on images of other datasets. Further, external clinical validation was performed by comparison of the algorithmic segmented images with those generated by a manual segmentation done by an experienced radiologist. We have termed our network as CCN-PR-Seg-net.