TCPNet: A Novel Tumor Contour Prediction Network Using MRIs

Brain cancer ranks among the top ten causes of death globally and stands as the second leading cause of cancer-related deaths among adolescents. Magnetic Resonance Imaging (MRI) is widely used as one of the most important diagnostic imaging techniques for early detection of tumors in brain cancer. In general, the contour is manually identified by experts from the MRIs, causing errors due to subjective assessments. In order to address this issue, image segmentation techniques can be used to automate the process of identifying and delineating the tumor region from the MR images. In this work, we propose a novel deep-learning framework named TCPNet, which is developed in the spirit of the U-Net model. The proposed architecture ensures that the model segments the tumor contours and explicitly estimates data and model uncertainties in the predictions, which is essential for tumor contour detection. Our extensive study on two open-source brain MRI datasets shows that TCPNet performs better than U-Net and other state-of-the-arts in terms of common evaluation metrics. Additionally, the proposed model presents uncertainties in model predictions, demonstrating confidence in segmented tissues or advising for expert intervention if necessary.