Leveraging Uncertainty Estimates to Improve Segmentation Performance in Cardiac MR
Résumé
In medical image segmentation, several studies have used Bayesian neural networks to segment and quantify the uncertainty of the images. These studies show that there might be an increased epistemic uncertainty in areas where there are semantically and visually challenging pixels. The uncertain areas of the image can be of a great interest as they can possibly indicate the regions of incorrect segmentation. To leverage the uncertainty information, we propose a segmentation model that incorporates the uncertainty into its learning process. Firstly, we generate the uncertainty estimate (sample variance) using Monte-Carlo dropout during training. Then we incorporate it into the loss function to improve the segmentation accuracy and probability calibration. The proposed method is validated on the publicly available EMIDEC MIC-CAI 2020 dataset that mainly focuses on segmentation of healthy and infarcted myocardium. Our method achieves the state of the art results outperforming the top ranked methods of the challenge. The experimental results show that adding the uncertainty information to the loss function improves the segmentation results by enhancing the geometrical and clinical segmentation metrics of both the scar and myocardium. These improvements are particularly significant at the visually challenging and difficult images which have higher epistemic uncertainty. The proposed system also produces more calibrated probabilities.
Origine : Fichiers produits par l'(les) auteur(s)