- Abstract: Automatic detection of tumorous tissue in MRI scans plays an important role in computer-aided diagnosis. We present a novel deep fully convolutional encoding architecture for semantic segmentation of brain MRI scans termed, FR-MRInet. This trainable encoder works with a corresponding decoder of a fully connected network. The 32 layer deep encoding architecture is inspired by VGG16 and InceptionV3. The novelty of FR-MRInet is its architectural design that efﬁciently reduces input to a lower resolution feature map(s). The encoder uses strides instead of pooling in various layers to reduce feature maps without loosing spacial information. We used a non-overlapping sliding window with and a novel activation function called, Relu-RGB to train the model so that the model directly produces the ﬁnal output instead of a mask. We compared our model with well known imagenets such as Alexnet and VGGnet, other recent models proposed by researchers testing for pixel wise accuracy,intersection over union(IoU)and mean square loss value. We conducted our experiment on BRATS dataset for benchmarking and one of the latest dataset which was proposed in 2016 consisting of, 3064T1-weighted contrast-enhanced images from 233 patients. We also show that FR-MRINet provides an impressive performance on live images detecting tumors as well. To further investigate the matter, we have consulted with an MD about the usefulness and the future of these kind of projects. Our code is open sourced and freely available at github.com/xxxxxxxxx/FR-MRInet.
- Keywords: Deep Convolutional Neural Networks, Semantic Pixel-Wise Segmentation, Encoder, MRI scan, Brain Tumor