Uncertainty-Aware Brain Tumor Segmentation in MRI: A Res-UNet Framework with Monte Carlo Dropout

Precise identification of tumors in the brainis crucial for clinical diagnosis as well as for treatment planning. However, tumor heterogeneity, its irregular boundaries and poor availability of labeled data have impeded automated segmentation models' efficacy and reliability. In this paper, we present a deep learning architecture framework that leverages the high representational power of Residual U-Net (Res-UNet) and the uncertainty quantification with MonteCarlo dropout. The Res-UNet architecture enhancestheprediction accuracy of the model by adding residual connections to the model. This has the effect of solving the well known problem of vanishing gradientsand enabling deep feature representations. We use Monte Carlo sampling at inference with dropout layers to meet the clinical expectation for reliablepredictions. This approach allows the modelto measure its own segmentation outputs and epistemic uncertainty. The $U$ map gives crucial information to help in clinicaldecision making by pointing where exactly the model is not too sure. The performance of the model is verified with thehelp of our good Tversky index81 of 91.68% when we tested our method on the LGG Kaggle Brain MRI dataset. Our segmentation results becamemore clinically useful and interpretable by incorporating uncertainty quantification and our framework was showed to be a very promising AI-based solution in neurooncology.