MRI segmentation using deep learning network for brain tumour detection

Gliomas are a combination of infiltrating tumour cells and vasogenic edema. The abscission and radiation intensified in this region will improve survival. It is difficult to distinguish infiltrating cells with conventional imaging sequences. This paper presents an accurate and automatic method for defining areas of tumour infiltration in peritumoral edema in brain MRI, using a fully convolutional neural network, employing semantic segmentation technique. The architecture has a contracting path capturing the features and a symmetric expanding path enabling precise localisation similar to U-Net. The expansive path yields a U-shaped architecture. The multiparametric pattern analysis from clinical MRI sequences assists in identifying the tumour recurrence in peritumoral edema. This helps resection and strengthening of postoperative radiation therapy. In the proposed model, dice similarity coefficient metric (0.99, 0.98, 0.98) for complete, core and enhancing regions are obtained. Positive predictive value and sensitivity of corresponding regions are (0.98, 0.98, 0.98).